UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Classification of windthrow on outblock boundaries from Landsat 7 ETM Ortlepp, Stephanie Maren 2003

Your browser doesn't seem to have a PDF viewer, please download the PDF to view this item.

Item Metadata

Download

Media
831-ubc_2004-0044.pdf [ 12.06MB ]
Metadata
JSON: 831-1.0075082.json
JSON-LD: 831-1.0075082-ld.json
RDF/XML (Pretty): 831-1.0075082-rdf.xml
RDF/JSON: 831-1.0075082-rdf.json
Turtle: 831-1.0075082-turtle.txt
N-Triples: 831-1.0075082-rdf-ntriples.txt
Original Record: 831-1.0075082-source.json
Full Text
831-1.0075082-fulltext.txt
Citation
831-1.0075082.ris

Full Text

CLASSIFICATION OF W I N D T H R O W O N C U T B L O C K B O U N D A R I E S F R O M L A N D S A T 7 E T M by STEPHANIE M A R E N O R T L E P P B.S.F, The University of British Columbia, 1999 A THESIS SUBMITTED IN P A R T I A L F U L F I L M E N T OF T H E REQUIREMENTS FOR T H E D E G R E E OF M A S T E R OF SCIENCE I N FORESTRY in T H E F A C U L T Y OF G R A D U A T E STUDIES T H E F A C U L T Y OF FORESTRY Department of Forest Resources Management We accept this thesis as conforming to the required standard T H E UNIVERSITY OF BRITISH C O L U M B I A October 2003 © Stephanie Maren Ortlepp, 2003 Library Authorization In presenting this thesis in partial fulfillment of the requirements for an advanced degree at the University of British Columbia, I agree that the Library shall make it freely available for reference and study. I further agree that permission for extensive copying of this thesis for scholarly purposes may be granted by the head of my department or by his or her representatives. It is understood that copying or publication of this thesis for financial gain shall not be allowed without my written permission. Stephanie Ortlepp 19/01/2004 Name of Author (please print) Date (dd/mm/yyyy) Title of Thesis: Classification of Windthrow on Cutblock Boundaries from Landsat 7 ETM Degree: MSc. Department of Forest Resources Management The University of British Columbia Vancouver, BC Canada Year: 2004 ABSTRACT The use of Landsat 7 ETM satellite data to detect windthrow on cutblock boundaries on North Vancouver Island was investigated in this study. Mature western hemlock (Tsuga heterophylla (Raf.) Sarg.), Pacific silver fir (Abies amdbilis (Dougl.) Forbes), and Sitka spruce (Picea sitchensis (Bong.) Carr) stands in that area are susceptible to windthrow due to the stand structure and topographic exposure. A n endemic windthrow problem has arisen on cutblock boundaries, road edges, and riparian strips. Two approaches were used to predict the windthrow on the cutblock boundaries. The first used logistic regression analysis to model probability values for the pixels in the Landsat scene. The second used supervised classification to identify the windthrow pixels. The input data for both these methods included the six multispectral Landsat 7 ETM bands, and a series of forty vegetation indices that were calculated for the image data. The results indicate that using only the logistic regression models to identify windthrow pixels is insufficient. However, the supervised classification showed a 72.7% success rate in finding high severity windthrow pixels on cutblock edges. The highest accuracy in detection was for large areas of heavy windthrow. i i TABLE OF CONTENTS ABSTRACT i i LIST OF TABLES v LIST OF FIGURES v i ACKNOWLEDGEMENTS vii 1 INTRODUCTION 1 1.1 Objectives 3 2 LITERATURE REVIEW 4 2.1 Windthrow and Remote Sensing 4 2.2 Logistic Regression Analysis 6 2.3 Supervised Classification 8 2.3.1 Minimum Distance 9 2.3.2 Maximum Likelihood 11 2.4 Variables Affecting Classification Results 12 2.5- Vegetation Indices 13 3 METHODS 15 3.1 Study area 15 3.2 Field data 17 3.2.1 Phase 1: Reconnaissance 19 3.2.2 Phase 2: Ground Truth Data 19 3.3 Remotely Sensed Data 22 3.3.1 Data Pre-processing 22 3.4 Modeling Windthrow Areas 27 3.4.1 Logistic Regression Analysis 27 3.4.1.1 Two-Class Model 27 3.4.1.2 Three-Class Model 31 3.4.2 Supervised classification 32 3.4.2.1 Two-Class Model 32 3.4.2.2 Three-Class Model 34 4 RESULTS 35 4.1 Extent of Windthrow in Study Area 35 4.2 Modeling windthrow areas 35 4.2.1 Logistic Models Based on Two Classes for Windthrow Area 35 4.2.2 Logistic Models Based on Three Classes for Windthrow Severity 41 4.2.3 Supervised classification 47 ii i 5 DISCUSSION 59 5.1 Logistic Models 59 5.2 Supervised Classification 61 5.3 Alternative Approaches to Detecting Windtrhow 65 6 CONCLUSIONS 67 LITERATURE CITED 69 APPENDIX I REPORT O N SPECTRAL LINEAR UNMIXING 76 iv LIST OF TABLES Table 1. Landsat 7 E T M and IKONOS bands 23 Table 2. Number of training area polygons and pixels for each of the training areas 26 Table 3. List of vegetation indices 28 Table 4. Number of pixels in each model building set and test set 29 Table 5. List of input channels used in the supervised classifications 33 Table 6. Summary of small plot volumes by species and damage level 36 Table 7. Summary of the two-class logistic models 40 Table 8. Accuracy estimates for the two-class logistic models when tested on the test sets 42 Table 9. Accuracy estimates for the two-class models when tested on the full data sets 43 Table 10. Summary of the three-class logistic models 46 Table 11. Accuracy estimates for the success of three supervised classifications in correctly identifying small plots as windthrow 51 Table 12. Summary of the percent of supervised classifications which correctly identified each plot as windthrow 58 v LIST OF FIGURES Figure 1. Comparison of the minimum distance and maximum likelihood classifiers 10 Figure 2. Stand types found in the study area 16 Figure 3. Location of the study area 18 Figure 4. Landsat 7 E T M panchromatic image of the study area showing the block and plot placements 20 Figure 5. Training area locations and colour designations 25 Figure 6. Example of a Class one plot with a low percent windthrow by volume 37 Figure 7. Example of a Class two plot with a medium percent windthrow by volume 38 Figure 8. Example of a Class three plot with a high percent windthrow by volume 39 Figure 9. Graphs of windthrow volume of the small plots versus the probability values for a) the two-class model 1; and b) the two-class model 4 44 Figure 10. Probability levels of windthrow for the study area using a) the two-class Model 1, and b) the two-class Model 4 45 Figure 11. Scatter plots showing a) high correlation between band 3 and band 2; and b) low correlation between band 4 and band 3 48 Figure 12. Supervised classification using only the six multispectral bands with the maximum likelihood classifier 49 Figure 13. Percentage of the large and small ground plots identified as windthrow for the supervised classifications using band and ratio combinations selected from the two-class logistic regression models 52 Figure 14. Percentage of the large and small ground plots identified as windthrow for the supervised classifications using band, ratio, and probability image combinations from the two-class logistic regression models 53 Figure 15. Supervised classification using the band, ratio, and probability level combinations from model four, with the maximum likelihood classification method 55 Figure 16. Supervised classification accuracies for all cover types as calculated from self classification of the training areas 56 vi ACKNOWLEDGEMENTS This thesis would not be in its present form without the help and support of many people. I would first like to thank my parents for their years of moral and financial support. During my many years in university, they encouraged me through the more difficult times and shared my triumphs and joys as if they were their own. I am also deeply grateful to my committee members: Peter Murtha, for teaching me so many things, for his encouragement, support and advice, and for always taking an interest. Val LeMay, for patiently explaining statistics to me - several times, braving the first draft of my thesis, and being a friend. Steve Mitchell, for showing me my first area of windthrow in Port McNeill, and giving much needed advice on the field-work, without which it would have taken at least twice as long. I am grateful the many people and friends at UBC, who were always willing to give advice and help whenever I needed it. There are simply too many people to thank, but I would like to especially mention Jerry Maedel, and the people of the FIRMS lab, who made my time as a Master's student a joy. This thesis would not have been possible without the funding from Forest Renewal BC and the Forestry Innovation Investment Ltd., and the very generous contributions of Western Forest Products in Port McNeill. vii 1 INTRODUCTION Windthrow is a natural part of the forest ecology. Although it has not received the same attention as other forest health issues such as mountain pine beetle, it none-the-less causes considerable loss of timber volume. Windthrow occurs throughout the province of British Columbia (BC), causing thousands of hectares of timber to be lost annually in uncut stands, and along boundaries of cutblocks and roads (Stathers et al. 1994). In 1992, the BC Ministry of Forests conducted a windthrow survey and found that the damaged volume was equal to 4% of the allowable annual cut, which at the time, was comparable to the damage caused by fire or insects (Mitchell 1995). Besides the loss of volume, windthrow can also lead to other forest health issues. Increased fuel loading can result in high fire hazards (Stathers et al. 1994). Also, stressed and uprooted trees in areas of windthrow can provide excellent conditions for bark beetles, which can then spread to the surrounding healthy trees (Stathers et al. 1994; Mitchell 1995; Peltonen 1999). In natural forests, windthrow can have beneficial effects on the forest and soils, including gap creation to allow for advanced and new regeneration, and mixing of mineral and organic layers to create favourable seedbeds and maintain soil fertility (Brown 1977; Ruel 1995). However, under current forest practices, a problem has been created on cutblock boundaries, road edges, and riparian strips, where trees are suddenly exposed to winds that they are not capable of withstanding. Once windthrow has occurred, the decision must be made whether it is economic to salvage the windthrown trees. This requires forest managers to have up-to-date and accurate knowledge of the amount and location of windthrow. Furthermore, automatic detection of windthrow from remotely sensed images would assist in the development of empirical windthrow risk models (Mitchell et al. 2000). Many factors affect the severity and location of windthrow. Tree characteristics, which determine the wind-firmness of individual trees, include: tree height, diameter, bole 1 taper, bole strength and elasticity, shape and size of crown, rooting factors such as depth, area, size and density of roots, and interlocking of roots. Trees with large crowns, low stem taper, and a shallow root plate are most susceptible to windthrow. Stand level characteristics also play an important role. Stand height and density, species composition, and any silvicultural treatments that may have been done, interact to make a stand more or less wind-firm (Stathers et al. 1994). Soil characteristics, such as drainage, and structure, determine how well the tree roots can hold against the forces created by winds. Soil moisture, as well as the rain or snow load on tree crowns, factor in as well. The wind exposure, direction, and turbulence, and meteorological conditions such as wind speeds, gustiness, and duration all affect the location and severity of windthrow (Stathers et al. 1994). Windthrow can be classed into two levels, depending on the severity of damage and frequency of damaging winds. Catastrophic windthrow is caused by unusually strong winds which only occur sporadically over a long time period, and cause local damage to extensive patches. Endemic windthrow occurs more frequently, but produces smaller patches. Additionally, wind damage types can be classed as: 1) stem break, when the tree snaps at some height above ground level; 2) stock break, when the tree snaps at ground level; 3) root break, when the tree falls due to failure of the roots just below ground level; or 4) tree throw of hinge fall, when the whole tree is uprooted (Stathers et al. 1994). Currently, windthrow management deals mostly with preventative measures such as cutblock orientation and shape, feathering edges, and topping trees (Stathers et al. 1994). Windthrow monitoring of stands is not routinely done. The Windthrow Handbook (Stathers et al. 1994) outlines procedures which involve ground surveys, and the use of large scale maps to annually update areas having windthrow. These methods do not provide information frequently enough to allow preventative measures to be taken, nor can they be easily integrated into windthrow models. There are several advantages that digital satellite data have over ground-based monitoring. Satellite data are synoptic, multi-temporal, and multi-spectral, the analysis 2 of the digital data is objective and repeatable, and data can easily be integrated with a GIS (Geographic Information System). Since satellite data also cover a much broader area and are available in a range of resolutions, both spectral and spatial, analysis can be done at the stand or landscape-level. In an area where windthrow has occurred, there wil l be a different surface exposure of the wood, bark and soil, and a lower percentage of green foliage. This creates spectral differences between windthrow areas and areas with standing trees, or cutblocks with slash (Jackson et al. 2000). These differences should be detectable using remote sensing methods, once a windthrow threshold has been reached. 1.1 OBJECTIVES The objectives of this study were: 1. To compare the use of logistic regression models with supervised classification, in the identification of windthrow areas using only spectral reflectance as input; 2. To determine the level of windthrow severity which can be detected using Landsat 7 ETM image data; and 3. To determine the minimum area of windthrow that can be detected with Landsat 7 ETM image data. In order to address these objectives, Landsat 7 ETM data from North Vancouver Island were obtained on October 01, 2001, and used in conjunction with ground truth data to build a series of logistic regression models and supervised classifications which were applied to the image data. The models were built using two different datasets: a two-class dataset (windthrow or no windthrow) for windthrow area, and a three-class dataset (high severity, low severity, and no windthrow) for windthrow severity. It was hypothesized that most of the cover types would be well differentiated from the windthrow class with higher success using the three-class models than for the two-class models in both the logistic regression models and the supervised classifications. 3 2 LITERATURE REVIEW 2.1 W I N D T H R O W A N D R E M O T E SENSING Remote sensing is both a science and an art that deals primarily with the spectral reflectance signatures of surface features. The underlying concept is that each surface feature has a unique spectral signature, and as the surface of the Earth changes, so do the spectral signatures. Thus, when a forest canopy is changed by an event such as windthrow, the new surface features can be picked up by satellite imagery. Initially, when windthrow occurs, there are significant changes to the canopy, understory vegetation, and soil surface. The uprooting causes soil disturbance, greater bark exposure for logs, and foliage drying. The spectral signature that is found in a new area of windthrow is unique, because it displays a certain proportion of bark, dead/dying foliage, and exposed soils. This unique collection of signatures would allow the distinction between areas of new windthrow and other cover types such as full canopy or cutblocks. Over time, new vegetation takes root, and understory vegetation and advanced regeneration, not killed by the windthrown trees, grows with new vigour. After several years, the ground may be completely covered by vegetation, masking all the formerly exposed soil, and possibly covering the boles of the windthrown trees. As the area subjected to windthrow passes through successional stages, so does the spectral signature of that area. The dynamics of windthrow have been extensively studied (Blackburn et al. 1988; Galinski 1989; Helliwell 1989; Stathers et al. 1994). However, the actual problem of monitoring windthrow has not received as much attention. Only a handful of studies have used satellite data to study windthrown gaps in a forest. Landsat digital data have been proven to be effective in monitoring changes in forest clearings (Sader 1995; Collins and Woodcock 1996), and were used by Nelson et al. (1994) to study large patches of windthrow (greater than 31 ha) in the Brazilian Amazon forests. They found that the use 4 of Landsat T M data quite successfully distinguished between areas of windthrow versus anthropogenic clearings and secondary forests. A German study by Kahabka et al. (2001b) also found that 70% of windthrow patches were found by visual interpretation of pan-enhanced Landsat ETM data. The same study found that the higher resolution IRS-1C satellite allowed for the detection of areas as small as 0.5 ha with 87% accuracy. A similar study conducted by Ramminger et al. (2001) found that radar data were effective in finding areas of windthrow larger that 2 hectares. Murtha (2000a; 2000b; 2000c; 2001) used RADARSAT satellite C-band, Fine 2 beam-mode radar data to monitor windthrow in riparian leave strips on northern Vancouver Island. Multi-temporal data were used to measure holes caused by windthrow in 60 riparian strips. Windthrow along the edge of cutblocks was also recorded during the study, as was the relationship between the level of windthrow and the landform. Murtha found that the highest levels of windthrow correlated strongly with two landforms, namely, till over rock and ground moraine. Satellite-based imagery for monitoring has several advantages over the use of ground or aerial surveys including: 1. Multispectral imagery gathers data over a broader wavelength spectrum; 2. Data can be used for many other purposes beside windthrow monitoring; 3. Multi-temporal data can be used to monitor change; 4. Data are readily available and cheaper than ground surveys; 5. Digital analysis techniques, such as the use of vegetation indices, supervised classification, and spectral linear unmixing, can be used to improve forest damage assessment; 6. Satellite data are easy to update; and 7. Digital, georeferenced-monitoring data can be readily integrated with other landscape level datasets using available GIS software. However, the use of satellite imagery has not yet been incorporated into standard management practices in BC. 5 2.2 LOGISTIC REGRESSION A N A L Y S I S Logistic regression analysis is used to explore the relationships between discrete responses and a set of explanatory variables. The discrete responses can be either in the form of binary responses (Yes or No for the presence of windthrow), or ordinal responses (High density, low density, or no windthrow). Linear logistic regression is used to model discrete responses, and is fit using methods for maximum likelihood (SAS Institute Inc. 1999; Smith et al. 2002). For a binary windthrow response the logistic regression model is represented by the following equations: (1) /(*) = ln = J30+ /3xx, + p2x2 + ...+ pkxk +E (2) 1 + e where: f(x) is the function modeling the odds ratios p is the probability of windthrow /?o is the intercept parameter xi through %k are explanatory variables in the form of bands or vegetation indices /?i through [3k are the slope parameters E is the error term The results of the logistic analysis include the estimates of the j9 parameters for each significant explanatory variable, and a series of diagnostic measures that indicate the quality of the model. The ordinal windthrow logistic regression utilizes the proportional odds model. The model is based on the cumulative probabilities of the response categories, rather than their individual probabilities; otherwise, the equations are essentially identical to the binary windthrow model. For an ordinal windthrow model with three classes (i.e., high density windthrow, low density windthrow, and no windthrow), the probability values 6 from the model are for: a) high density windthrow; b) high + low density windthrow; and c) all three classes with probability of 1.0. This allows the calculation of individual probabilities for each of the classes. A pixel is placed into the class to which it has the highest probability of belonging. The following equations are used to obtain the cumulative probability values (Hosmer and Lemeshow 2000): (3) / f a H n (4) /(XJH 1-A (5) p= 6 { P1+P2 ^ |=A>(2) l + efM where: f(xt )are the functions modeling the cumulative odds ratios V is the probability of windthrow /?0 ( i ) are i intercept parameters Xi through Xk are k explanatory variables in the form of bands or vegetation indices f3i through are the k slope parameters E1 and E 2 are error terms Logistic regression models have been used with remotely sensed data in several studies to model the relationships between a set of response and explanatory variables. For example, Koutsias and Kartens (1998) used logistic regression modeling of Landsat T M data for burned area mapping. Another study was done on forest fire ignition probability, where it was found that estimated accuracies of over 90% could be attained using the logistic models (Perestrello de Vasconcelos et al. 2001). Smith et al. (2002) used logistic regression to model the effects of patch size and land cover heterogeneity on the success of image classification, and found that both negatively affected the accuracy that could be realized. Several studies have focused on building predictive models with logistic analysis, including studies on urbanization (Gunter et al. 2000), on deforestation in Belize (Chomitz and Gray 1996), and a study on land cover change in Cameroon (Mertens and Lambin 2000). A study by Ludeke et al. (1990), built a methodology for 7 using logistic modeling to predict anthropogenic deforestation in tropic countries, and found that a high rate of success (87%) could be attained with relatively few variables. 2.3 SUPERVISED C L A S S I F I C A T I O N Supervised classification is an analysis tool that classifies each pixel in an image into one of a set of user-defined classes. A training area is a group of pixels which represent an individual cover type, such as windthrow. Each cover type present in an image must receive its own training area, which is selected based on a priori knowledge. The training areas are used to build the model which allows for the classification of each pixel in the image; therefore, it is important to have training sites which are representative of the cover type. Scatter plots can be used to determine some of the characteristics of the training sites. They are graphical representations of a training site spectral signature over two bands, and show the level of correlation between the two bands (Lillesand and Kiefer 2000). Low correlation is indicated by a plot with well scattered cloud of points, whereas high correlation is indicated by the points showing a distinct linear distribution along the 45 degree line (Lillesand and Kiefer 2000). Low correlation is more desirable because it indicates that each of the bands contains unique information, which would allow a better separation between the cover classes. In the classification stage, each pixel is placed into the cover class to which it comes closest based on the training sites and the classification method used. Supervised classification has received extensive application in resource monitoring (Lillesand and Kiefer 2000; Muchoney and Strahler 2002; Smith et al. 2002). For example, Hudak (1993) found that using supervised classification on Landsat T M data successfully classified defoliation of balsam fir (Abies balsamea (L.) Mill.) by the hemlock looper (Lambdina fiscellaria fiscellaria (Guen.)). A 2001 study in Germany exploring the use of satellite data (including IKONOS, IRS-1C, SPOT 2, and Landsat ETM) to monitor catastrophic wind events on forests found that IKONOS data provided good results with the use of supervised classification (Kahabka et al. 2001a; Kahabka et al. 2001b). Other 8 examples include Foody and Hi l l (1996), who used classification to separate tropical forest classes, Knick et al. (1997), who used image classification on semi-arid rangeland, and Fiorella and Ripple (1993), who classified successional stages in temperate coniferous forests. Both the choice of variables (Bands or Channels) and classification method directly affect the classification results. Two classification methods that are commonly used are minimum distance and maximum likelihood. 2.3.1 Minimum Distance The minimum distance classifier uses the Euclidean distance calculated between the spectral signature of each pixel and the mean vector of all pixels (the average spectral value in each of the bands) of each tiaining site class. Each pixel is then assigned to the class based on the minimum Euclidean distance. This classifier uses the following equation (PCI 2002): (6) Gc<P>$U?pi-Mj where: Gc(p) = spectral distance for pixel p to the mean class c n = number of channels in classification i = a particular channel c = a particular class ]id = mean digital number for channel i for all pixels of class c xPi = digital number of pixel p in channel i The minimum distance approach results in each pixel being assigned a class; however, it does not necessarily provide the most accurate classification results. A pixel could be assigned to a class to which it doesn't belong, because the minimum distance does not take into account the variability of pixels within the training site clusters (Figure 1). The advantage of this classifier is that it is very fast (Lillesand and Kiefer 2000). 9 255 9 1 5 Distribution of Trairiing Site Pixels V. />— • V . /»—« '<-!>« >« Slash 4 Forest Type 1 Water Road Forest Type 2 Band 4 Digital Number 255 2 5 5 Maximum Likelihood Classifier 2 5 5 21 5 c Minimum Distance Classifier =2 Band 4 Digital Number Band 4 Digital Number 255 Figure 1. Comparison of the minimum distance and maximum likelihood classifiers using a two-band example. The " X " represent unclassified pixels, which are assigned to the training area identified in green, according to the method of classification being represented. Source: adapted from Lillesand and Kiefer, 2002 10 2.3.2 Maximum Likelihood The maximum likelihood classification method is one of the most commonly used algorithms, due to its high classification accuracy. While the rruhimurn distance classifier uses averages in calculating the Euclidean distance, this classifier uses both the means and variance/co-variances of the training class spectral response patterns. For each training class, the mean vector and a variance-covariance matrix for all spectral bands are calculated. When the classifier is applied to the image, each pixel's spectral signature and the class statistics are used to calculate the probability of each class belonging to that pixel. The pixel is then assigned to the class to which it has the highest probability of membership (Figure 1). This classifier assumes multivariate normality, namely that the distribution of points making up each training class of each input channel is Gaussian. If a training site shows a non-Gaussian, multi-modal distribution, then it is representing two or more sub-classes (Lillesand and Kiefer 2000). The log likelihood for a multivariate normal distribution is calculated as follows (PCI 2002): (7) / ^^- l lo^n^lo^J- i^ -z / jV^^-^^log^) (8) / / c = t » c » A " l (9) x p = \ x ] p x 2 p x l . . . x n p ] where: lc{p) - log-likelihood for pixel p to belong to class c Vc = the (n by n) matrix with the variance and covariance values between different channels for the cth class V~l = the inverse variance-covariance matrix for the c"1 class |Vcj = the determinant of the covariance matrix for class c xp - the (n by 1) vector of D N values for the pth pixel, where xp is the channel 1 D N value for the pth pixel n = the number of channels in the classification JUC = the (n by 1) mean vector for class c, where p] is the channel 1 mean for class c Pc = the a priori probability for class c 11 The first part of the equation is a constant and does not affect the classification outcome. The second part is the determinant of the variance-covariance matrix, which is a measure of the area of the hyper-ellipsoid in the feature space, and, in essence, penalizes classes that have a higher variability. The third part is the Mahalanobis distance, which accounts for the different variability of clusters. The inverse of the variance-covariance matrix defines the shape and orientation of the hyper-ellipsoid in the feature space, and the final term in the equation is the a priori probability for a class, which defaults to 1.0, meaning that all classes have the same probability of being selected. This term can be used to modify the classifier so that classes which are known to cover more area in an image receive a higher weight during the classification (Schowengerdt 1983). 2.4 V A R I A B L E S A F F E C T I N G CLASSIFICATION RESULTS There are several natural variables that can affect the success of a classification. The topography can affect the scene radiance depending on the slope and aspect of an area. This necessitates having several training sites for one class, or having several classes for one cover type. The sun angle affects how much shadow there is, thus on the edge of a cutblock, this can cause difficulties in detecting the windthrow. In areas of shadow little spectral reflectance is picked up by the satellite sensor; thus, areas of windthrow that are on the shaded (south) edge of a cutblock will be difficult or impossible to detect. The view angle also has an impact, since a surface material will reflect differently when viewed from different angles. This becomes significant when working with an image that covers a large area (such as a full Landsat scene), since the view angle changes from nadir to off-nadir at the edges of the image (Schowengerdt 1983). Finally, patch size or higher heterogeneity of a surface cover also makes classification more difficult since the training sites contain so much variability that they may start to overlap in the feature space (Smith et al. 2002). The classification of a pixel as windthrow wil l depend on a complex interaction of variables. The percent of windthrow and the original stand density and structure will affect the spectral reflectance. The percent canopy loss will also determine how much of the ground will be visible to the satellite sensor. A l l of these affect the severity threshold at which a pixel will be identified as windthrow. 12 2.5 V E G E T A T I O N INDICES Since the spectral information in an image is represented by Digital Numbers (DNs), these can be used in mathematical functions. Vegetation indices (Vis) are simply mathematical transformations of some or all of the available multispectral data, and fall into one of two categories. The ratio or slope-based indices are simple ratios of bands or band combinations. They are linear combinations that represent the slopes of the spectral reflectance curves of the bands that make up the ratio. They show the spectral characteristics of an image while eliminating the variations in illumination caused by topography or aspect. Thus, slope-based indices can help to identify subtle variations in the spectral reflectance that would normally be masked by brightness variations (Elvidge and Chen 1995). A n example of this type of index is the Normalized Difference Vegetation Index (NDVI) (Pearson and Miller 1972): NDVI=^^-(10) NIR + R The NDVI is a ratio of the near-infrared (NIR) and red (R) wavelengths, and it is one of the most frequently used indices for vegetation analysis (Lillesand and Kiefer 2000). Orthogonal Indices are the other type of VI. Since individual bands of a multi-spectral image are often highly correlated, they appear similar and convey essentially the same information. Principal component analysis (PCA) is commonly used to remove or reduce redundancy in multispectral data. PCA results in a linear transformation that produces a weighted linear combination of some or all of the original image bands. PCA reduces the dimensionality (the number of bands) of the original data, and compresses information (or variance) from the original data into the least number of new components. A n example of this type of index would be the Perpendicular Vegetation Index (PVI) (Richardson and Wiegand 1977), which contains only the red and near-infrared bands multiplied by factors derived from PCA, to allow maximum differentiation between vegetation types: (11) PVI = -0.S13R + 0AS66NIR 13 Many studies have been done on the success of a variety of vegetation indices to find relationships between the spectral reflectance and ground conditions. For example, several satellites, including Landsat TM, were used with a series of ratios to estimate leaf area index (LAI) and percent green cover (Elvidge and Chen 1995; Purevdroj et al. 1998). Another study by Asrar et al. (1984) modeling the LAI and photosynthetically active radiation (PAR) found that use of the NDVI allowed estimation of the PAR (±10%) and LAI (±0.8units) from measures of spectral reflectance. A series of ratios were used to estimate the green biomass in semi-arid grasslands by Anderson et al. (1993). Allen and Kupfer (2001) used the Tasseled Cap transformations in a study on the spectral response of Fraser fir (Abies fraseri (Pursh) Poir.). Vegetation indices have also been used to model leaf water stress (Cohen 1991b), assess fire hazard (Cohen 1991a), classify vegetation based on photosynthetic activity (Lloyd 1990) , model ecological variables such as soil and meteorological conditions (Cihlar et al. 1991) , and model coniferous forest characteristics (McDonald et al. 1998). 14 3 M E T H O D S 3.1 STUDY A R E A The study area is located on northern Vancouver Island, BC, surrounded by the communities of Port Hardy, Port McNeill and Port Alice, on Western Forest Products, Ltd., and Weyerhaeuser Ltd. tree farm licenses. It is located in the Nahwitti Lowlands Ecosection and in the very wet maritime subzone of the Coastal Western Hemlock (CWH) Biogeoclimatic Zone (Krajina 1965; Klinka et al. 1991). The dominant tree species are western hemlock (Tsuga heterophylla (Raf.) Sarg.), western redcedar (Thuja plicata Donn), Pacific silver fir (Abies amabilis (Dougl.) Forbes) and Sitka spruce (Picea sitchensis (Bong.) Carr) (Lewis 1985). There are two main forest types in the study area (Weetman et al. 1990)(Figure 2): 1. C H type, which is old growth forest dominated by western redcedar, western hemlock, and a small component of Pacific silver fir. These stands are open and multistoried, with a salal (Gaultheria shallon Pursh.) dominated understory; and 2. H A type. This type is dominated by western hemlock and Pacific silver fir. These stands are second growth, dense, even-aged, with a sparse understory. According to Lewis (1982), this stand type arose from a series of windthrow events, the last of which occurred in 1906. The geomorphology of this area has been shaped by a series of glaciations, the last of which occurred approximately 14,000 years ago (Mollard and Janes 1984). As a result of this, the surface geological materials are predominantly composed of unconsolidated morainal and fluvial outwash overlying bedrock. The landforms that predominate in the study area are ground moraine over basal till and till over rock, both of which have been associated with high susceptibility to windthrow (Murtha 2001). Also found in the study area are marine 15 Figure 2. Stand types found in the study area. The left side of the image shows the hemlock-balsam (HA) stand type on the well drained slopes, which is highly susceptible to windthrow. The right side shows the cedar (CH) stand type on the poorly drained base of the slope, which is much more wmd-firm. 16 and lacustrine deposits. These are poorly drained and support the more open C H types, which are less susceptible to windthrow. The C W H very wet maritime subzone is characterized by having mild winters and cool moist summers, with a mean annual precipitation of about 1700 mm, and mean annual temperatures of 7.9°C. The predorninant type of natural disturbance in the area is windthrow (Prescott 1996), most of which occurs during the fall and winter. During this period, coastal British Columbia is subjected to mid-latitude cyclonic storms, which are associated with high winds (Sieben 2001). Northern Vancouver Island has a history of catastrophic and endemic windthrow events (Weetman et al. 1990). These stands are now mature and susceptible to windthrow again (Stathers et al. 1994). Already endemic windthrow damage from winter winds along recently exposed cutblock edges, ranges from individual trees to patches several hectares in size (Mitchell et al. 2000). The group of cutblocks that was selected for study were located between O'Connor Lake and Rupert Inlet (Figure 3), part of an area that has undergone extensive study as part of the Salal Cedar Hemlock Integrated Research Program (SCHIRP) (Prescott 1996) and clearcut edge windthrow risk modeling (Mitchell et al. 2000). The study was restricted to the H A stand types, as they display many characteristics that make them high risk for windthrow. These stands are fairly dense, with trees that are tall, have a low stem taper, and relatively shallow rooting. 3.2 FIELD D A T A Ground truth data were collected in two phases. First, a reconnaissance trip was used to establish the overall amount of windthrow, and second, data on the windthrown and residual standing volumes were collected. 17 Figure 3. Location of the study area on North Vancouver Island. 1 8 3.2.1 Phase 1: Reconna issance The boundary of every cutblock in the study area was surveyed, and the condition of each 10m segment of the boundary was documented. For each segment, the following information was gathered: 1. Amount of windthrow by severity of canopy loss (Low: 0-30%, Medium: 31-60%, High: 60%+); 2. Distance of penetration from the edge of the cutblock was estimated to the nearest metre; 3. Notes on any relevant attributes, such as species composition and age of windthrow (fresh vs. old); and 4. Photographs were taken of each cutblock, and of the largest areas of windthrow. In total, 24 cutblocks were visited, of which three were eliminated (two were actively being logged, and one was over 15 years old), leaving 21 cutblocks for the study (Figure 4). 3.2.2 Phase 2: G r o u n d T r u t h D a t a Two different types of fixed-area plots were established. The model-building plots were 30 by 30-metres, to correspond with one Landsat pixel. These were carefully placed in the largest windthrow areas found along the cutblock edges, as they were intended to be used as the windthrow tiaining areas for building the models. These plots were distributed to provide variation in the density of the original H A stand, the level of windthrow, and the height of regeneration and ground vegetation. In total, six of these plots were established (Figure 4). The second type of plot was 10 by 30-metres, and oriented parallel to the original stand edge along the cutblock boundaries. Based on the reconnaissance, each area of windthrow that was longer than 30 metres was assigned a number, 35 of which were randomly selected for the plot locations. These plots were intended to be used to test the logistic models and supervised classifications in their ability to identify windthrow pixels. Additionally, 19 Figure 4. Landsat 7 ETM panchromatic image of the study area showing the block and plot placements. The cutblocks making up the study area are outlined in blue, and numbered in black. The red stars and numbers mark the locations of the small plots, and the yellow stars mark the location of the large plots, identified by the cutblock number. Plots 49, 50, 88, 98, and 101 are in areas of heavy shadow, and were not used to test the supervised classifications. 20 boundary segments, in which no windthrow was found, were noted, to use as non-windthrow test areas. For both plot types, the information collected was: 1. The global position system (GPS) location of the plot (corrected using the Burnaby BC Reference Station) to the nearest metre; 2. A photograph (Canon PowerShot A-20 digital camera) of sub-canopy ground cover, and a vertical photograph of the underside of the canopy from the middle of the plot to provide an estimate of the canopy closure; and 3. A sketch of the location of each tree and record of ground cover types. For each tree which had initiated in the plot, the following were collected: 1. Diameter outside bark at breast height (DBH; 1.3m above ground) using a diameter tape (d-tape) to the nearest 0.1 cm; 2. Diameter at top (0 if not snapped) using d-tape to the nearest 0.1 cm; 3. Length of the tree (height if standing) using Irnpulse-2 laser hypsometer to the nearest 0.1 cm; 4. Species; and 5. Condition (windthrown, leaning/broken, standing/dead). The final step for the field data collection was a helicopter fly-over to take aerial pictures of each plot, to aid in the determination of ground cover types and percentages, and the amount of canopy cover. The volume of each tree of all field plots were calculated using the species volume equations from the Forestry Handbook for British Columbia (Watts 1983). The small plot data were summarized by plot and species, to obtain the volumes of windthrown and standing trees. The volumes for the leaning trees were calculated using the measured D B H and estimates of the tree height made in the field. For snapped trees, attempts were made to find the top sections of the tree to get a complete length, otherwise the height of the tree was estimated from the surrounding trees of similar diameter. 21 3.3 R E M O T E L Y SENSED D A T A Data from Landsat 7 ETM were acquired on October 01, 2001 (Table 1). A n IKONOS scene was acquired for August 11, 2001. Additionally, ArcMap (ESRI Inc. version 8.3) vector overlays of the road networks, TFL boundaries, and cutblock boundaries, and hardcopy maps of the area were provided by Western Forest Products, Port McNeill, BC. 3.3.1 Data Pre-processing Sub-scenes covering the extent of the study were extracted from the Landsat scene. The Landsat and IKONOS scenes were re-projected to N A D 83 datum and Universal Transverse Mercator (UTM) projection using the Focus and Xpace modules of PCI Geomatica version 8.2.1 (PCI 2002) with the nearest neighbor re-sampling algorithm. This algorithm rotates the image data to the new coordinate system without averaging any of the pixel values, but has a maximum spatial error of half the pixel resolution. The vector overlays were also re-projected to the same datum and projection. The IKONOS scene was registered to the Landsat scenes. Additionally, the IKONOS scene was overlaid on the Landsat scene by rotating it in Adobe Photoshop version 7.0.1 (Adobe Systems Inc., version 7.0.1). By using Photoshop, the pixels were re-oriented and the values were not re-sampled, allowing the original IKONOS pixels to be used when later locating the ground plots. Ground plots were geo-registered with the Landsat data by using a combination of the GPS data, vector data, and the rotated IKONOS image. Once the plots were located on the Landsat scene, the pixel values in the form of digital numbers for six multispectral bands were extracted using the Xpace module of PCI Geomatica, and were entered into an Excel 2000 (Microsoft Corp., version 9.0.4402) spreadsheet. Because the Landsat image was acquired in the morning, five of the 35 small plots fell into areas of heavy shadow, and were of no use in further analysis. Due to the plot sizes and orientations, plot boundaries did not correspond to the Landsat pixel boundaries. A l l Landsat pixels that crossed the plot 22 Table 1. Landsat 7 ETM and IKONOS bands. Landsat 7 ETM IKONOS Bands Wavelength Pixel Resolution Bands Wavelength Pixel Resolution B l Blue 450-515 nm 30 m Blue 450-520 nm 4 m B2 Green 525-605 nm 30 m Green 510-600 nm 4 m B3 Red 630-690 nm 30 m Red 630-700 nm 4 m B4 Near Infrared (NIR) 775-900 nm 30 m NIR 760-850 nm 4 m B5 Short-wave Infrared 1 (SWIR1) 1.55-1.75um 30 m B6« Thermal Infrared 10.4-12.5um 60 m B7 Short-wave Infrared 2 (SWIR2) 2.09-2.35um 30 m B8« Panchromatic 520-900 nm 15 m Panchromatic 450-900 nm 1 m Source: (Lillesand and Kiefer 2000) a Landsat bands not used in analysis 23 boundary were used for subsequent analysis, and a weighted average of the pixel values was calculated using the percent of each pixel that overlapped with the ground plot. A l l field data were entered into an Excel 2000 spreadsheet in order to calculate basic plot information on the windthrow levels and species distributions. Once all ground and Landsat data were geo-registered, twenty four training areas were digitized onto the 2001 Landsat scene (Figure 5). These areas represented the different cover types that were characteristic of the study area, and they were placed in large, uniform areas of the surface feature that they were to represent. The non-windthrow training areas were selected based on the higher-resolution IKONOS data and ground truthing done during field visits. Some of the large ground plots were used to place the training areas representing windthrow. Where possible, each training area was represented by two or more polygons distributed over the extent of the study area (Table 2). According to Hixson et al. (1980), all cover types must be adequately represented by a sufficient number of pixels in order to guarantee a true representation of the spectral characteristics of the classes. Schowengerdt (1983) also notes that a sufficient number of pixels must be used to estimate the training class spectral characteristics accurately. This is especially important for the maximum likelihood classifier where the assumption of multivariate normality is made. In order to calculate the class mean vectors and variance-covariance matrices, each class must contain at least n+1 pixels, where n is the number of input channels. If fewer than n pixels make up a training class, the inverse of the covariance matrix cannot be calculated as its determinant is zero. This, however, is the theoretical minimum, and in practice, at least lOn to lOOn pixels should be in each training class (Swain and Davis 1978). Additionally, as the within-class variability increases, more pixels are needed to accurately calculate the class signature characteristics (Schowengerdt 1983; Story and Campbell 1986). A final consideration in regards to the training areas is the number of land cover types represented by framing classes. When there are too few classes, the feature spaces are so broad that they overlap, resulting in many misclassification errors. If there are too many classes, difficulties arise in matching the spectral classes with the surface cover types in the image (Richardson et al 1976). 24 Bog L_J BrightRegerYl Regenl L _ Shade forest cm " 1 BrightRegen2 Regen2 Shaded Cut CH2 1 BrightRegen3 Regen3 t Shadow = = GravelPit !• NewCutl 1 1 Regen4 Water-Lake Hfll ] NewCut2 1 1 Regen5 i Water-Ocean Z] HA2 n NewCuO !__] TidaFlats c • Windthrowl Windthrow2 q Figure 5. Training area locations and colour designations. 2 5 Table 2. Number of training area polygons and pixels for each of the fraining areas used in the model building process for the logistic regressions and supervised classifications. Training Area Number of Total Number Name polygons of pixels Bog 1 27 CH1« 3 57 CH2 1 34 Gravel Pit 1 11 HAV 2 118 HA2 1 38 Bright Regen lc 1 166 Bright Regen 2 2 21 Bright Regen 3 2 15 New Cut 1 2 56 New Cut 2 2 36 New Cut 3 1 43 Regen 1 2 31 Regen 2 1 42 Regen 3 1 62 Regen 4 1 33 Regen 5 3 67 Tidal Flats 1 70 Shaded Forest 1 36 Shaded Cut 1 24 Shadow 2 28 Water-Lake 3 102 Water-Ocean 1 189 Windthrow ld 2 12 Windthrow 2d 2 14 a C H = western redcedar dominated stand types. b H A = western hemlock - Pacific silver fir dominated stand types. c Regen = regeneration. d For the 2-class models, the two windthrow classes were merged to form a single windthrow class. 26 The DNs were extracted for each training area pixel for the six multispectral bands using the Xpace module of PCI Geomatica, and were entered into an Excel 2000 spreadsheet. Forty vegetation indices (Vis) were selected based on their successful use in other studies, or were created for the purpose of this study, since little is known on the success of Vis in the detection of windthrow (Table 3). The Vis were calculated using the DNs for each of the small plot pixels, the weighted average of the small plot pixels, and each of the training area pixels in Excel 2000. 3.4 M O D E L I N G W I N D T H R O W A R E A S 3.4.1 Logistic Regression Analysis The logistic models relate the digital spectra with the on-ground conditions, and were built and tested according to the methodology developed by Snee (1977). This method involves splitting the data set into a model-building and a test set. The model building data were used to establish the relationship between the area or severity of windthrow and the spectral reflectance. The test set was set aside and used to test the fitted models. The models were built using the DNs and Vis for the six multispectral bands of the twenty four training areas. The data for each training area were randomly divided into four equally sized groups. These groups were then separated into four different model-building and test sets, where the model set consisted of 75% of the data, and the test set was the remaining 25% (Table 4). Using this method, models were built and tested for each of four different datasets. Two sets of models were built using these methods: the first set was built using two classes, to allow the estimation of windthrow area; and the second method used three classes to estimate the severity of windthrow. 3.4.1.1 Two-Class Model For two-class set of models, all non-windthrow framing areas were labeled class 0, whereas the windthrow training areas were labeled class 1. This set of models would be used to indicate the area of windthrow. 27 Table 3. List of vegetation indices used in the models and applied to the Landsat 7 ETM scene. B l to B7 represent the multispectral bands described in Table 1. Label RI R2 R3 R4 R5 R6 R7 R8 R9 RIO R l l R12 R13 R14 R15 Ratio B4-B3 B4+B3 Normalized Difference Vegetation Index (Pearson and Miller 1972) B4-B3 B4+B3 Ratio Vegetation Index Gordon 1969) B4 53 B4 Ml B3 B3 B2 B2\ B4 B3 B4 B4 B2 B2-B3 B2+B3 B4-B2 B4+B2 B4\ B2-B3 B2+B3 B3 BI+B2+B3 B2 BI+B2+B3 Bl BI+B2+B3 Label R16 R17 R18 R19 R20 R21 R22 R23 R24 R25 R26 R27 R28 Ratio B4 B2+B3+B4 B3 B2+B3+B4 B2 B2 + B3 + B4 Transform Vegetation Index 1 (Rouse et al. 1973) B l B1+B2+B3 +0.5 Transform Vegetation Index 2 (Rouse et al. 1973) JNDVI+0.5 Perpendicular Vegetation Index (Richardson and Wiegand 1977) (-0.873x£3>HP.4866xB4) B5-B4 B5+B4 B1-B5 B1+B5 B3 Bl £5 B2 B5 B4 Bl B3 BJ_ B5 " R33 = 0.1544S1+0.2552B2+0.3592B3+0.5494B4+0.549B5+0.4228B7 fcR34= -0.1009B1-0.1255B2-0.2866B3+0.8226B4-0.2458B5-0.3936F7 CR35= 0.3191Bl+0.5061fi2+0.553453+0.0301B4-0.5167B5-0.2604B7 Label R29 R30 R31 R32 R33 R34 R35 R36 R37 R38 R39 R40 Ratio B1-B4 B1+B4 B5 B4+B5+B1 Bl B4+B5+B1 B4 B4+B5+B1 Tasseled Cap Brightness (PCI 2002)° Tasseled Cap Greenness (PCI 2002)fc Tasseled Cap Wetness (PCI 2002)c B1-B5 B1+B5 +0.5 B5-B4 B5+B4 +0.5 B1-B4 B1+B4 +0.5 B5-B1 B2 B2 Soil Adjusted Vegetation Index (Huete 1989) B4-B3 1.5 B4+B3+0.5 28 Table 4. Number of pixels in each model building set and test set for both the Two-Class and Three-Class models. Training Area Name Total Model 1 Model 2 Model 3 Model 4 Number Model Test Model Test Model Test Model Test of pixels Set Set Set Set Set Set Set Set Bog 27 20 7 20 7 20 7 21 6 CH1« 57 42 15 43 14 43 14 43 14 CH2 34 25 9 26 8 26 8 25 9 Gravel Pit 11 8 3 9 2 8 3 8 3 HAV 118 88 30 88 30 89 29 89 29 H A 2 38 29 9 29 9 28 10 28 10 Bright Regen l c 166 125 41 125 41 124 42 124 42 Bright Regen 2 15 11 5 11 5 11 5 12 4 Bright Regen 3 21 16 4 16 4 16 4 15 5 New Cut 1 56 42 14 42 14 42 14 42 14 New Cut 2 36 27 9 27 9 27 9 27 9 New Cut 3 20 32 11 33 10 32 11 32 11 Regen 1 31 23 8 24 7 23 8 23 8 Regen 2 42 32 10 31 11 31 11 32 10 Regen 3 62 46 16 47 15 47 15 46 16 Regen 4 33 25 8 25 8 24 9 25 8 Regen 5 67 50 17 51 16 50 17 50 17 Tidal Flats 70 52 18 53 17 52 18 53 17 Shaded Forest 36 27 9 27 9 27 9 27 9 Shaded Cut 24 18 6 18 6 18 6 18 6 Shadow 28 21 7 21 7 21 7 21 7 Water-Lake 102 77 25 76 26 76 26 77 25 Water-Ocean 189 142 47 142 47 142 47 141 48 Windthrow l d 12 9 3 9 3 9 3 9 3 Windthrow 2d 14 11 3 10 4 10 4 11 3 Total 1332 998 334 1003 329 996 336 999 333 " C H = western redcedar dominated stand types. b H A = western hemlock - Pacific silver fir dominated stand types. c Regen = regeneration. d For the 2-class models, the two windthrow classes were merged to form a single windthrow class. 29 The training area data, consisting of the DNs for the six multispectral bands and the forty ratios calculated from these DNs, were imported into SAS (SAS Institute Inc. version 8.02), and a stepwise logistic procedure was used to relate the variables to windthrow class. From the stepwise logistic procedure, a list of the significant (a=0.05) multispectral bands and/or ratios was obtained. The variables and their estimated parameters were used to predict the probability that a given pixel was windthrow. This was repeated for the remaining three model-building sets. To assess the model fit, a series of diagnostic measures were obtained for each model. These include the maximum R 2 value, the -2Log Likelihood (-2LogL), and the Akaike Information Criterion (AIC). The maximum R 2 value is an indication of the spread of points around the model. The -21ogL can be used to test whether all regression coefficients in the model are simultaneously zero, or if at least one of the regression coefficients is non-zero. The AIC penalizes the -21ogL value for including more predictor variables, and is most useful when comparing several models. Better models have a maximum R 2 value close to one, and low -21ogL and AIC values (Hosmer and Lemeshow 2000; Smith et al. 2002). The equations from each model-building set could be applied to any pixel in the Landsat scene, producing a probability value, ranging from 0 to 1, that the given pixel was windthrow. Error classification tables were produced for each model, by testing on the 25% of the original data that had been set aside as the validation set, and by testing on the full dataset. A cutoff probability value was selected for each model. Pixels that had a value above the cutoff would be identified as windthrow and pixels below the cutoff would be identified as non-windthrow. The cutoff value was chosen to maximize the correct identification of windthrow pixels, and minimize the error of commission. The logistic models were applied to the Landsat scene of the study area, producing windthrow-probability-level images for each model. The observed were compared to expected results based on the IKONOS images, stand-type maps, and the large and small plot data. It was expected that large areas of windthrow would have the highest probability values, and areas of non-windthrow would have low probability values. 30 However, it was also expected that some of the cover-types (e.g., cedar stand types, some of the older cutblock classes, and road edges) would receive higher probability values and be indistinguishable from windthrow due to the similarity of their spectral reflectance curves to that of windthrow. 3.4.1.2 Three-Class Model For the three-class models, all non-windthrow areas were labeled class 0, and the windthrow haining areas were divided into low initial stand volume with 100% windthrow (class 1) and high initial stand volume with 100% windthrow (class 2). This set of models would be used to indicate the severity (in terms of initial stand density) of windthrow. This was done in an attempt to alleviate the multi-modal distribution of the windthrow training class. The procedure to build the three-class models was essentially the same as for the two-class model set. The stepwise logistic procedure was used to relate the variables to windthrow class. From the stepwise logistic procedure, a list of the significant multispectral bands and/or ratios (cr=0.05) was obtained. The variables and their parameters were used to predict the probability of belonging to the severity classes of windthrow of each pixel. Error classification tables were produced for each model, by testing on the 25% of the original data that had been set aside as the validation set, and by testing on the full dataset. The logistic models were applied to the small plot data to test their ability on a wider range of windthrow severities. It was expected that plots with a higher percent windthrow by volume would produce a higher probability value. The observed results were compared to the predicted results by graphing the percent windthrow by volume of the small plots to their probability value produced by the model. The cumulative probabilities of windthrow (i.e., combined probabilities for class 1 and class 2) were also used to produce an image of the probability of each pixel being windthrow for the study area. It was expected that the results would be better than those of the two-class model sets. 31 3.4.2 Supe rv i sed c lass i f i ca t ion The same training areas from the logistic regression analysis were used in the supervised classifications. The Focus module of PCI Geomatica version 8.2.1 was used to perform the analysis. Both the minimum distance (MD) and maximum likelihood (ML) methods were used on each set of input variables. The input variables for both the two-class and three-class classifications were selected from the six multispectral bands (Bl to BT), the 40 Vis (RI to R40), and the probability level images ( M l to Ma) (Table 5). 3.4.2.1 Two-Class Model The supervised classifications using the two-class training areas was based on the training dataset with one windthrow class, and 23 non-windthrow training classes. The first classification used only the multispectral bands as input variables, to provide a base-line classification. The next five classifications used the bands and ratios selected for the four logistic models built on the two-class dataset. The next six classifications used the six multispectral bands in combination with the probability level imaged from the two-class logistic models. The final classifications were again based on the variables identified by the four logistic models, but also included the probability level images. The classification success of the non-windthrow areas was estimated by using the overall accuracy obtained from the error matrix of the training site classification. Self-classification of the training set data is somewhat biased since, the training sites are usually the best representations of a surface cover type, whereas other parts of the image with the same cover type will be less homogenous. However, this method gives a rough idea of the classification accuracy that can be realized for the entire image (Lillesand and Kiefer 2000). The image classification accuracy was considered important since the context of the pixels identified as windthrow is important, and the information from the classification could be applied to other uses as well. The observed and expected results were compared based on the IKONOS images, stand-type maps, and the large and small plot data. It was expected that most of the cover types would be well differentiated from the windthrow class. However, the roads, the western redcedar stand types, and some of the cutblock classes would likely be mixed 32 Table 5. List of input channels used in the supervised classifications. The label column identifies the classification attempt. B l to B7 represent the six multispectral bands, R3 to R39 represents the vegetation indices from, M l to M4 represent the probability-level images from the two-class model, and Ma represents the average of these four probability images. The x's identify which input channels were used in each of the classifications. B6 is the thermal band with lower resolution and is, therefore, not used in the analysis. Label B l B2 B3 B4 B5 B7 R3 R4 R6 R7 R8 R13 R18 R23 R24 R26 R29 R30 R33 R37 R38 R39 M l M2 M3 M4 M a 1 X X X X X X 2 X X X X X X 3 X X X X 4 X X X X X 5 X X X X X X 6 X X X X X X X X X X X X X X 7 X X X X X X X 8 X X X X X X X 9 X X X X X X X 10 X X X X X X X 11 X X X X X X X 12 X X X X X X X X X X 13 X X X X X X X 14 X X X X X X X X 15 X X X X X X X X X X 16 X X X X X 17 X X X X X X 18 X X X X X X X X 19 X X X X X X 20 X X X X X X X 21 X X X X X X X X X 22 X X X X X X 23 X X X X X X X 24 X X X X X X X X X X 25 X X X X X X X X X X X X X X X 1 X X X X X X 2 X X X X X X 3 4 X X X X X X X 5 X X X X X X 6 X X X X X X X X X X X X X 7 X X X X X X X 8 X X X X X X X 9 X X X X X X X 10 X X X X X X X 11 X X X X X X X 12 X X X X X X X X X X 13 X X X X X X X X 14 X X X X X X X 15 X X X X X X X X X 16 X X X X X X 17 X X X X X X X X 18 X X X X X X X X X X 19 X X X X X X X X 20 X X X X X X X X X 21 X X X X X X X X X X X 22 X X X X X X 23 X X X X X X X 24 X X X X X X X X X X 25 X X X X X X X X X X X X X X 60 60 O - J 01 -C *-t g % TJ o » s to o o H rs U TJ Ol a. 3 Cft C _o "5 01 ca 60 O hJ Ol 43 Ol C _ O TJ TJ O 2 0 01 Ol Cft Ol .3 H v U TJ Ol 0) eft 33 with the windthrow class. It was also expected that the plots with both higher relative and absolute windthrow volumes would be classified as windthrow more consistently, whereas the plots with lower relative and absolute windthrow volumes would likely not be classified as windthrow. 3.4.2.2 Three-Class Model This set of classifications was based on the training dataset with two windthrow classes (one for high- and one for low-density windthrow), and 23 non-windthrow fraining classes. This set of classifications followed the same pattern as the two-class set. The probability level images used for these classifications were those from the two-class logistic regression models, as it was felt that the two-class images would better represent the presence of windthrow. The classification success of this set was also estimated by using the overall accuracy obtained from the error matrix of the training site classification. The observed and expected results were again compared based on the IKONOS images, stand-type maps, and the large and small plot data. It was expected that most of the cover types would be well differentiated from the windthrow class with higher success than for the classifications using the two-class model. It was also expected that the plots would be classified with higher success than the two-class model. 34 4 RESULTS 4.1 E X T E N T OF W I N D T H R O W I N STUDY A R E A The small plot data were summarized by plot, species, windthrow percent and volume, and standing tree percent and volume (Table 6). The average windthrow over the small plots was 44.5% by volume, of which 31.3 % was western hemlock (H), 11.2 % Pacific silver fir (B), 0.8% western redcedar (C), and 1.2 % sitka spruce (S). Additionally, of all the H trees in the small plots, 37.9% of them were windthrown, of the B, 61.4% were windthrown, of the C, 41.4% were windthrown, and of the S, 25.4% were windthrown. This shows that for the cutblocks surveyed, although H is the dominant species, B is much more susceptible to windthrow. The percent windthrow by volume was used to group the small and large plots by windthrow severity: Low-Class 1 (<30%), Medium-Class 2 (31%-60%), or High-Class 3 (60+%) (Figure 6 to 8). Class 1 contained 10 small plots and no large plots, Class 2 contained 14 small plots and one large plot, and Class 3 contained six small plots and five large plots. 4.2 M O D E L I N G W I N D T H R O W A R E A S 4.2.1 Log is t i c M o d e l s Based o n T w o Classes for W i n d t h r o w A r e a The first set of four models was produced from the dataset containing one windthrow class and one non-windthrow class. Table 7 summarizes the parameters and variables for each model resulting from the stepwise logistic procedure, and the diagnostic measures which indicate the model success. Some of the variables were repeated in several of the models, such as vegetation index R37, which was found in models 1, 3, and 4. Other variables that were repeated in two or more models were: B2, R4, R18, R23, and R30. The models had maximum R 2 values ranging from 0.47 to 0.59, -2LogL values ranging from 85.21 to 103.76, and AIC values ranging from 99.21 to 113.76. 35 Table 6. Summary of small plot volumes by species and damage level. The windthrow category includes all trees that displayed stem break, stock break, root break, or hinge fall, and windthrow Class 1 = <30%, Class 2 = 31-60%, and Class 3 = 60+% windthrow by volume Block Plot No. Standinc volume by species (m3/ha) Standing Volume all Species (m3/ha) Windthrow volume by species (rrfVha) Windthrow Volume all Species (m3/ha) Total Plot Volume (m3/ha) Percent Windthrow by Volume Windthrow Class Ha B c S H B c S 220 1 494.47 97.19 0 0 591.66 144.72 292.58 0 0 437.30 1028.96 42.5 2 221 5 489.77 28.05 0 0 517.82 105.79 132.55 0 0 238.34 756.16 31.5 2 7 437.60 0 0 0 437.60 183.18 99.00 0 0 282.18 719.78 39.2 2 223 9 7.10 0 0 0 7.10 423.37 524.92 0 0 948.29 955.39 99.3 3 10 692.20 258.77 0 0 950.97 21.79 372.40 0 c 394.19 1345.16 29.3 1 224 16 96.35 0 0 0 96.35 640.04 0 0 0 640.04 736.39 86.S 3 17 1036.27 39.69 0 0 1075.96 11.77 0 0 c 11.77 1087.73 1.1 1 19 1147.40 56.21 0 0 1203.61 67.14 110.57 0 0 177.71 1381.32 12.S 1 20 741.20 0 0 0 741.20 457.97 0 0 0 457.97 1199.17 38.2 2 225 22 691.91 0 0 0 691.91 50.26 0 0 0 50.26 742.17 6.8 1 23 622.64 56.07I 0 0 678.71 36.61 385.43 0 0 422.03 1100.74 38.3 2 226 26 25.76 513.26 0 0 539.02 123.84 216.72 c 0 340.56 879.58 38.7 2 27 643.31 0 49.55 0 692.85 223.87 0 c 0 223.87 916.72 24.4 1 229 31 1079.57 0 0 0 1079.57 77.01 0 c 0 77.01 1156.58 6.7 1 34 q 0 C 0 0 319.26 0 c 0 319.26 319.26 100.C 3 35 700.65 170.20 167.35 0 1038.20 413.57 259.01 c 0 672.58 1710.78 39.3 2 562 40 19.03 0 0 0 19.03 296.32 43.80 0 0 340.13 359.16 94.7 3 42 720.36 542.83 0 0 1263.19 155.65 0 0 193.97 349.61 1612.80 21.7 1 44 1172.48 0 16.93 127.14 1316.54 57.94 d 0 0 57.94 1374.48 4.2 1 564 49° 358.68 0 0 0 358.68 255.29 C 0 63.05 318.33 677.01 47.C 2 50° 799.42 0 9.26 0 808.67 361.08 C 3.07 0 364.15 1172.82 31 .C 2 52 57.21 0 0 0 57.21 598.14 420.99 0 c 1019.13 1076.34 94.7 3 574 58 5.84 720.31 0 0 726.16 226.86 751.8C 0 0 978.66 1704.82 57.4 2 578 61 574.39 C 0 0 574.39 448.41 55.28 0 c 503.69 1078.07 46.7 2 65 411.72 185.06 0 0 596.78 592.15 0 c c 592.15 1188.93 49.8 2 74 0 C 0 0 0 614.13 404.29 c 0 1018.42 1018.42 100.C 3 594 80 151.38 C 0 0 151.38 1053.85 134.72 c 0 1188.57 1339.94 88.7 3 81 172.69 c 0 436.81 609.50 290.87 117.0G c 52.69 460.65 1070.16 43.C 2 5307 85 717.89 c 0 274.92 992.82 898.19 0 c 74.06 972.25 1965.06 49.5 2 88° 520.56 c 62.35 420.71 1003.62 83.39 0 97.99 106.49 287.87 1291.49 22.3 1 91 1216.18 c 0 54.19 1270.36 296.56 0 C 0 296.56 1566.93 18.S 1 5310 98° 236.92 c 0 0 236.92 819.46 0 35.48 38.39 893.33 1130.25 79.C 3 5313 99 351.22 c 91.11 281.60 723.93 448.73 0 175.76 13.90 638.39 1362.32 46.S 2 100 554.50 c 16.94 0 571.44 345.13 0 59.74 0 404.86 976.30 41.5 2 5409 101° 997.48 140.89 0 0 1138.36 108.5S 0 C 0 108.59 1246.95 8.7 1 H=western hemlock, B=Pacific silver fir, C=western redcedar, S=sitka spruce; b p ot eliminated because it is in heavy shadow 36 Figure 6. Example of a class one plot with a low percent windthrow by volume (Table 6). This image shows Plot 42 in cutblock 562. 3 7 Figure 8. Example of a Class three plot with a high percent windthrow by volume (Table 6). This image shows Plot 65 in cutblock 578. 39 Table 7. Summary of the two-class logistic models. Model Fitted Equation" Max. R2 -2LogL AIC No. Pixels6 1 f(x)= 291.10 - 3.69B3 - 57.48R4 + 0.97R8 - 476.10R18 - 141.80R30 + 76.31R37 0.55 92.76 106.76 998 2 f(x)= 36.58 - 1.07B2 - 9.00R24 + 31.88R30 - 11.70R38 0.47 103.76 113.76 1003 3 f(x)= 624.00 - 1.27B2 + 27.74R3 - 247.20R6 - 376.90R7 + 31.82R37 0.52 94.17 106.17 996 4 f(x)= 178.00 - 12.85R4 - 354.70R13 - 288.10R18 - 28.39R23 - 94.40R26 + 161.80R37 0.59 85.21 99.21 999 eflx,) a piwindthrow) T - — - ; definition of variables found in Table 3. f(xi) the function modeling for the high density windthrow class; \ + eHXi) i(x2) is the function modeling for the high + low density windthrow classes. b Number of pixels that were used to build the models 40 The models were tested on the reserved test set, and by adjusting the cutoff value, it was possible to weight the model in favour of classifying pixels as windthrow events. This resulted in a higher accuracy in detecting windthrow (WT), but an increase in the error of commission (Table 8). Model 1 had the highest accuracy (38.5%), and Model 2 had the lowest accuracy (22.7%). When the models were applied to the full training area data set, the results were somewhat different (Table 9). Of the four models, Model 1 again had the highest accuracy in detecting windthrow, at 41.7%, followed closely by Model 4 at 37.7%. Models 2 and 3 had much lower accuracies. The probability values were plotted against the percent windthrow volumes for the small plots (Figure 9a and b) to determine if the higher volumes of windthrow were being detected with higher accuracy. There was little correlation between the probability value and the percent windthrow volume for the plots, as shown for Models 1 and 4. The four models were converted into probability level images (Figure 10a and b). Since these are 8-bit images, the gray values range from 0 (black) to 255 (white). The black areas represent a low probability of windthrow, whereas white areas represent areas that have a high probability of windthrow. Areas of windthrow should appear as white, while other cover types should be dark. 4.2.2 Log is t i c M o d e l s Based o n Three C lasses for W i n d t h r o w Sever i t y The second set of models was produced using the three-class data set, and the same procedure was followed as for the first set of models. These models produced a different set of explanatory variables from the first set (Table 10). The variables that repeated in two or more of the models were: B3, R6, R18, R37, and R39. The models had maximum R 2 values ranging from 0.46 to 0.54, -2LogL values ranging from 104.09 to 142.14, and AIC values ranging from 110.06 to 126.14, which were generally worse than for the two-class models. When these models were tested on the reserve dataset, and the full dataset, it was found that they performed extremely poorly. None of the pixels in the low-density class were correctly 41 Table 8. Accuracy estimates for the two-class logistic models when tested on the test sets. Values indicate numbers or percents of pixels. Actual % Actual WTa non-WT sum WT non-WT Model 1 cutoff = 0.16 Predicted WT 5 1 6 38.5% 0.3% Predicted non-WT 8 320 328 61.5% 99.7% Predicted sum 13 321 334 Model 2 cutoff = 0.6 Predicted WT 5 2 7 22.7% 0.7% Predicted non-WT 17 304 321 77.3% 99.3% Predicted sum 22 306 328 Model 3 cutoff = 0.08 Predicted WT 5 2 7 29.4% 0.6% Predicted non-WT 12 317 329 70.6% 99.4% Predicted sum 17 319 336 Model 4 cutoff = 0.13 Predicted WT 4 2 6 30.8% 0.6% Predicted non-WT 9 318 327 69.2% 99.4% Predicted sum 13 320 333 " WT = windthrow; Non-WT = non windthrow (all other training areas) 42 Table 9. Accuracy estimates for the two-class models when tested on the full data sets. Actual % Actual WTa non-WT sum WT non-WT Model 1 Cutoff = 0.16 T J CD WT 20 6 26 41.7% 0.5% Predicl non-WT 28 1278 1306 58.3% 99.5% Predicl sum 48 1284 1332 Model 2 Cutoff = 0.60 •o CD WT 19 7 26 19.0% 0.6% Predicl non-WT 81 1225 1306 81.0% 99.4% Predicl sum 100 1232 1332 Model 3 Cutoff = 0.08 T3 CD WT 19 7 26 26.4% 0.6% Predicl non-WT 53 1253 1306 73.6% 99.4% Predicl sum 72 1260 1332 Model 4 Cutoff = 0.13 T3 CD WT 20 6 26 37.7% 0.5% Predicl non-WT 33 1273 1306 62.3% 99.5% Predicl sum 53 1279 1332 a WT = windthrow; Non-WT = non windthrow (all other training areas) 43 20 40 60 Percent Windthrow Volume 80 100 20 40 60 Percent Windthrow Volume 80 100 Figure 9. Graphs of windthrow volume of the small plots versus the probability values for a) the two-class model 1; and b) the two-class model 4. The red lines represent the probability cutoff values for the models. 44 a) b) Figure 10. Probability levels of windthrow for the study area using a) the two-class Model 1, and b) the two-class Model 4, with cutblock boundaries and plot locations from Figure 4. The cutblocks are outlined in blue, the red points are small plots, and the yellow points are large plots. White areas have a high probability of windthrow, and black areas have a low probability of windthrow. 45 Table 10. Summary of the three-class logistic models. Model Fitted Equation" R 2 -2LogL AIC No. Pixels 6 1 f(xi)= 85.63 - 1.56B3 - 19.18R6 - 143.20R18 - 54.92R37 - 10.94R38 - 641.90R39 f(x2)= 86.70 - 1.56B3 - 19.18R6 - 143.20R18 - 54.92R37 - 10.94R38 - 641.90R39 0.46 126.14 142.14 998 2 f(xi)= -60.91 - 3.26B2 - 11.11R24 - 219.20R26 + 981.70R30 + 0.47R33 - 2301.70R39 f(X2)= -59.73 - 3.26B2 - 11.11R24 - 219.20R26 + 981.70R30 + 0.47R33 - 2301.70R39 0.54 104.09 120.09 1003 3 f(xi)= 563.90 - 1.52B2 + 44.35R3 - 244.00R6 - 353.70R7 - 42.46R29 + 94.65R37 - 964.4R39 f(x2)= 565.00 - 1.52B2 + 44.35R3 - 244.00R6 - 353.70R7 - 42.46R29 + 94.65R37 - 964.4R39 0.52 170.45 126.31 996 4 f(xi)= 336.40 - 4.15B3 - 6.70R4 + 1.06R8 - 538.6R18 - 224.2R30 + 106.6R37 f(x2)= 337.70 - 4.15B3 - 6.70R4 + 1.06R8 - 538.6R18 - 224.2R30 + 106.6R37 0.54 110.06 126.06 999 " p(windthrow) — -7--; definition of variables found in Table 3. f(xi) the function modeling for the high density windthrow class; l + e / u , ) f(x2) is the function modeling for the high + low density windthrow classes. b Number of pixels that were used to build the models 46 classified as such, and only 25% of the high density class pixels were correctly classified by the best of the models. Since these models performed so poorly, they were not converted into probability level images for use with the supervised classifications. The results from testing the models on the full data sets were similar. 4.2.3 Supervised classification The first step in the supervised classification process was to determine the quality of the training areas for supervised classification. The scatterplots showed that there were high correlations between only some of the bands, and much lower correlation between the different ratios, and the ratios and bands (Figure 11). The histograms for most of the classes showed a normal distribution, except for the windthrow class, which had a multi-modal distribution over all six multispectral bands. It was found that the classifications using the three-class dataset did very poorly in the overall accuracy for the non-windthrow classes, to the extent that the results were unrecognizable as the study area. Though, in a few cases, the windthrow pixels were being detected with some accuracy, it is also important to be able to place the windthrow areas in the context of the surrounding cover types and the whole image. For this reason, the results of the classifications using the three-class dataset will not be discussed in further detail. For the two-class set of classifications, using one windthrow training area, the first classifications were done using the M L method and only the six multispectral bands (Figure 12), to serve as a basis of comparison. When this classification attempt was compared to the IKONOS images, the classes that were accurately portrayed in the classification could be identified. The lake, ocean and new cutblock areas were classified with the highest success. The hemlock stand types and the various regeneration classes were also accurately portrayed on the classification; however, the cedar stand types, some of the road edges, and the shadow areas showed a strong association with the windthrow class. Most notably, the area identified as "A" in Figure 12 shows a classification that is a mix of shadow, cedar stand type and windthrow. This area was a river valley characterized by a mixed forest type, with frequent gaps in the canopy. The windthrow pixels identified at " B " are along another river. 47 a) 100 95 i 85 10 • 5 • 0 5 10 15 20 25 30 35 40 45 50 55 60 65 70 75 80 85 90 95 100 Band 2 Digital Number b) l l l i i : l I 1 1 1 1 1 1 I 1 1 I I I I 0 5 10 15 20 25 30 35 40 45 50 55 60 65 70 75 80 85 90 95 100 Band 3 Digital Number Figure 11. Scatter plots showing a) high correlation between band 3 and band 2; and b) low correlation between band 4 and band 3 (Table 1). 48 Figure 12. Supervised classification using only the six multispectral bands with the maximum likelihood classifier. The input channels used are from classification number 1 based on the two-class logistic regression models (Table 5). The colours represent the different surface cover types defined in the legend, and the original training areas used for the classification are outlined in black (Figure 5). The area identified as " A " is a river valley with a mixed forest type and frequent canopy gaps. The windthrow pixels identified at " B " are along the edge of a river. 49 The classification accuracies obtained in identifying all the small plots as windthrow showed fairly poor results. Classification 1ML had an accuracy of only 27.8% in correctly identifying the small plots as windthrow. The highest accuracy obtained was classification 24ML, which had an accuracy of 52.8% for identifying the small plots as windthrow, and the second best classification was 10MD which attained 50% (Table 11). Though these accuracies are not very encouraging, when the plots are split by windthrow class (Table 6), the results considerably improved for the high severity plots. The accuracies observed in detecting the large and small ground plots for the classifications using bands and vegetation indices as input channels, are summarized in Figure 13. This figure shows that 1ML correctly classified 55% of the high percentage windthrow (class 3) plots, 20% of the medium percentage windthrow (class 2) plots, and 40% of the low percentage windthrow (class 1) plots. Model 1ML had lower accuracies in classifying the class 2 and class 1 plots. Classifications 2MD (model 2 in Table 5 using minimum distance) and 2ML, 4MD, 5MD and ML, and 6MD, which included bands and vegetation indices identified from the logistic regression models, had higher accuracies in correctly classifying the large and small ground plots than 1MD. Classifications 3MD and 3ML, 4ML, and 6ML, however, had lower classification accuracies than 1MD. For the same set of input channels the input channels, in some cases, the minimum distance (MD) method performed better, and in others, the maximum likelihood (ML) method performed better, making it difficult to determine which method and model is best. Of this first set of the classifications, which did not contain the probability level images, the best results were from classification 5ML, followed closely by 5MD. Both of these classifications contained only the vegetation indices selected from the two-class logistic Model 4 (R4, R13, R18, R23, R26, and R37), which had the best diagnostic measures of the four models. The accuracies observed in detecting the large and small ground plots for the classifications which used the bands, vegetation indices, and probability level images from the four logistic regression models based on the 2-class dataset, as input channels are summarized in Figure 14. The first ten classifications have only the six multispectral 50 Table 11. Accuracy estimates for the success of three supervised classifications in correctly identifying small plots as windthrow. Actual % Actual WT3 non-WT sum WT non-WT I icted WT 10 1 11 27.8 % 1.8 % 1ML I icted non-WT 26 54 80 72.2 % 98.2 % Pred sum 36 55 91 70.3 % correct overall I icted WT 18 1 19 50.0% 1.8% 10MD I icted non-WT 18 54 72 50.0% 98.2 % Pred sum 36 55 91 79.1% correct overall I icted WT 19 2 21 52.8 % 3.6 % 24ML I icted non-WT 17 53 70 47.2 % 96.4 % Pred sum 36 55 91 79.1 % correct overall a WT = windthrow; Non-WT = non windthrow based no the 50 pixels selected from the field visit as being 100% non-windthrow. 51 140% o 120% A « 100% o 80% > 60% ** re I 40% 20% 0% • Windthrow Class 1 • Windthrow Class 2 • Windthrow Class 3 I I I I I I I I I I I I I I I I I I I I o CM Q CN - J Q ^ Q ^ Q CM CO m U5 Supervised Classification Label in to co Figure 13. Percentage of the large and small ground plots identified as wmdthrow for the supervised classifications using band and ratio combinations selected from the two-class logistic regression models. The supervised classification label identifies the input channels used in the classification from Table 5, and the classification method is identified as M D for minimurn distance and M L for maximum likelihood. The windthrow class is based on the percent windthrow by volume for the ground plots, where class 3 is high, class 2 is medium, and class 1 is low percent windthrow by volume (Table 6). 52 160% • Windthrow Class 1 •Windthrow Class 2 • Windthrow Class 3 0 = ! Q - ! 0 - ! 0 = i O f v . N - c 0 0 0 C T ) < : ' ^ O O T — r ' - T - ' " ^ ' - T - ' - ^ ' - ^ ' - T - ' - ^ ' - ( N t M N C M N ( V I N t \ l N t N ( M ( M Supervised Classification Label Figure 14. Percentage of the large and small ground plots identified as windthrow for the supervised classifications using band, ratio, and probability image combinations from the two-class logistic regression models as input channels. The supervised classification label identifies the input channels used in the classification from Table 5, and the classification method is identified as M D for minimum distance, and M L for maximum likelihood. The windthrow class is based on the percent windthrow by volume of the plots, where class 3 is high, class 2 is medium, and class 1 is low percent windthrow by volume (Table 6). 53 bands, and the probability level images as input channels. Of these ten classifications, 10MD (B1-B7, and the Model 4 probability level image) had a much higher accuracy than the other classifications. Additionally, this classification had a much higher accuracy than that of 1ML, as it identified 73% of the high windthrow (class 3) plots, 47% of the medium windthrow (class 2) plots, and 40% of the low windthrow (class 1) plots. The only classification that had a higher accuracy than 10MD was 24ML, which used the Model 4 vegetation indices, and all four probability level images. Classification 22MD, which also used the Model 4 vegetation indices, and only the Model 4 probability level image, showed identical accuracies to 10MD. The classification results from classification 24ML (Figure 15), overall, are quite similar to those of 1ML (Figure 12), with the exception that more pixels throughout the image are identified as the hemlock stand types and windthrow. Overall, this classification appears to be slightly more fragmented than 1ML. Additionally, the concentrations of windthrow pixels appear to be higher in certain areas, as shown by the windthrow pixels at "B" which are along a river. As in classification 1ML, the area identified as " A " is classified as a mixture of the shadow, cedar stand type and windthrow. The self-classification of the training sites showed an accuracy of 85% for 24ML, was which was only six percent lower than classification 1ML. The accuracies based on the self-classification of the training areas for all classification attempts are summarized in Figure 16. This accuracy measure averages the accuracies obtained for each cover type represented by a training area, and does not reflect that of windthrow alone. The training site classification accuracy for the classification using just the six multispectral bands and the maximum likelihood method (1ML; where 1 refers to the model numbers in Table 5) was 91%, which was the highest of all the classifications attempted. Most of the classifications showed a fairly high overall accuracy in this respect, and the M L method was better than the M D method in all cases, except for classifications 6MD and M L , where the M D method had a considerably higher accuracy. In order to compare the effects of various site variables on the classification of each plot, the percentage of classifications that successfully identified each plot as windthrow were 54 Bog BrightRegenl Regenl l _ J Shade forest 1 1 cm I I BrightRegen2 Regen2 Shaded Cut CH2 I I BrightRegen3 Regen3 Shadow GravePrl L_l NewCutl 1 1 Regen4 | Water-Lake • 1 HH1 n NewCut2 1 1 Regen5 | Water-Ocean HA2 i i NewCut3 1 1 TidalFlats | Windthrow • i Figure 15. Supervised classification using the band, ratio, and probability level combinations from model four, with the maximum likelihood classification method. The input channels used are from the two-class classification number 24 (Table 5). The colours represent the different surface cover types defined in the legend, and the original training areas used for the classification are outlined in black (Figure 5). 55 • MD 100% T • ML 90% Rn% . 1 1 _ 1 1 1 1 1 1 1 • • • • • _ _ • • _ • ou /o j >, 70% | o 2 3 60% | u < 50% j c g, 40% I a> °- 30% X 20% | 10% j 0% + 1 1 1 1 1 1 • I 1 1 1 1 1 1 1 1 1 1 1 1 1 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 Supervised Classification Label Figure 16. Supervised classification accuracies for all cover types as calculated from self classification of the training areas. The supervised classification label identifies the input channels used in the classification from Table 5, and the classification method is identified as M D for minimum distance, and M L for maximum likelihood. 56 compared to the percent canopy closure estimated from the field surveys, windthrow class from Table 6, and the amount of ground cover estimated from the aerial photos taken of each plot. The total number of classifications that correctly identified each plot as windthrow are summarized in Table 12. Typically, those plots with high windthrow (class 3), low canopy closure and low vegetation height were most consistently classified as windthrow. However, in some cases, even plots with high windthrow levels, and low canopy closure and low vegetation height were only identified by a few of the classification attempts (plot 74), while other plots with low volumes of windthrow, and high vegetation or canopy closure were identified by more of the classification attempts (plot 42). 57 Table 12. Surrvmary of the percent of supervised classifications which correctly identified each plot as windthrow. The windthrow class is based on the percent windthrow by volume from Table 6. The canopy closure is based on the canopy images taken in the field. The estimated ground cover is based on the aerial photos taken of each plot. The percent of classifications that correctly identified each plot as windthrow is based on both the M D and M L methods for the two-class model (Table 5). Plot No. Canopy Closure % Windthrow Amount of Percent of supervised classifications which Class Ground Cover a correctly identified the plot as windthrow 1 43.0 2 Low 4% 5 44.4 2 Low 0% 7 51.1 2 Low 0% 9 3.0 3 Low 84% 10 22.2 1 Med 98% 16 0.0 3 Med 92% 17 66.9 1 Low 24% 19 75.0 1 Low 0% 20 28.5 2 Low 0% 22 41.9 1 Med 0% 23 37.3 2 Low 40% 26 39.6 2 Low 6% 27 56.8 1 Low 18% 31 76.8 1 Low 38% 34 35.0 3 Med 4% 35 50.1 2 Low 46% 40 66.3 1 Low 52% 42 70.0 1 Low 72% 44 38.4 2 Low 34% 49 b 64.9 1 Low 0% 50* 8.7 3 High 0% 52 25.7 2 Low 82% 58 39.3 2 Med 74% 61 50.0 2 High 24% 65 0.0 3 Med 0% 74 0.0 3 Med 22% 80 32.5 2 High 0% 81 31.3 2 Low 48% 85 46.8 1 Low 30% 88* 50.4 1 Low 0% 91 4.9 3 Low 70% 98* 35.7 2 Low 0% 99 27.3 2 Med 4% 100 80.0 1 Low 6% 101* 67.6 1 Med 0% 223 3.0 3 Low 100% 229 30.0 3 Med 20% 594A 5.0 3 Med 96% 594B 7.0 3 Med 54% 5304 40.0 2 High 84% 5307 0.0 3 Low 98% "Low = little ground vegetation, exposed soil; Med = some vegetation covering soil and some boles; High = boles covered by vegetation; * plot eliminated because it is in heavy shadow 58 5 DISCUSSION 5.1 LOGISTIC M O D E L S There was little consistency in the variables that were found to be significant for the two-class logistic models among the four splits of the dataset. This would indicate that there is no single variable which is sensitive to windthrow alone. The maximum R 2 values were moderate, which results from the fact that windthrow is a rare event with an irregular spectral signature, and the few pixels that were used as haining areas could not explain the great deal of variation that would be found in areas of windthrow. This corresponds to the results by Eid and Oyen (2003), who found that mortality prediction in even-aged forests using logistic regression also yielded low R 2 values due to the relative infrequency of that event. The variables that were significant for two or more models were Band 2 (visible green wavelength), the ratio vegetation index (R4), R18, R23, R30, and R37. A l l of the Vis contained Bands 4, 5 or 7. Studies have indicated that the near infrared (Band 4) and the short wave infrared (Bands 5 and 7) regions are optimal at separating burned and unburned vegetation (Trigg and Flasse 2002). Additionally, Cohen (1991b) found that bands 5 and 7 showed the greatest change to different levels of leaf water stress. A study by Hardisky et al. (1983) also found that a vegetation index consisting of Bands 4 and 5 was very effective at detecting changes in plant biomass and water stress. Though the diagnostic measures were fairly similar for the models, on testing the models on the reserve and full data set, it was found that Model 1 had the highest accuracies by a considerable margin. When the models were tested on the plot data, it was found that there was only a weak relationship between the probability value obtained from the model, and the level of windthrow in the plot. Additionally, using the same probability cutoff values as for the test set, very few of the plots were identified as windthrow by 59 any of the models, meaning that the models cannot be used by themselves to predict the location of windthrow pixels. The probability level images showed that the large areas of windthrow received the highest probabilities; however, the cedar stand type, some sections of road, river valleys, and some of the older cutblocks also received higher values in some cases. The cedar stand types and the river valleys were characterized by a heterogeneous stand structure with gap openings. A l l these cover types are characterized by a component of exposed wood mixed with vegetation, which could indicate that the models depend on the presence of exposed wood to find the windthrow. However, several of the vegetation indices which made up the models appeared to be related to leaf moisture, which could indicate that this is also a contributing factor to the detection of windthrow. The logistic models based on the three classes were intended to predict the severity (in terms of originating stand density) of windthrow in the image. When these models were tested on the reserve sets, none of them were able to identify any of the low-density windthrow pixels and only one of the models identified any of the high-density windthrow pixels. When tested on the full dataset, the results were a little better for the high-density windthrow, though still none of the models identified the low density windthrow. This clearly indicates that these models cannot be used to predict the severity with a reasonable level of accuracy. The failure of these models to predict severity likely resulted from the fact that so few training pixels were available for either of the windthrow classes. As Koutsias and Karteris (1998) indicated in their study, the sample sizes for the two groups of data (e.g., wmdthrow, and non-windthrow for this study) should be about the same size, or the results would be biased. Unfortunately, it wasn't possible to use more wmdthrow pixels, as there were very few large windthrow sites available, and if smaller ones had been used, the quality of the training sites would have been compromised. 60 5.2 SUPERVISED C L A S S I F I C A T I O N It was found that due to the early morning acquisition time, all the cutblocks had at least one edge in shadow, which necessitated having several training classes for shadow. This also meant that any areas of windthrow along these edges would be undetectable. The only way to alleviate this problem would be to acquire images at noon in June; however, with satellites such as the Landsat series, this would be impossible as they have a fixed orbit. The time of year when an image is acquired would also affect the results of an analysis. In the winter, for example, snow cover would obscure many of the ground cover features that differentiate the windthrow areas from cutblocks, whereas in the fall, when the foliage of recently windthrown trees has dried out, classification may be easier. Cutblock edges are difficult to delineate on satellite data with 30 metre resolution, since they are the boundary between two distinct surface cover types. Windthrow occurs on this boundary, making it even more challenging to detect since its spectral signature wil l be mixed with that of two other surface cover types, unless it covers a large area. The lack of normality in the distribution of windthrow training class data was due to the fact that finding areas of windthrow large enough to use as training areas proved very difficult, and this necessitated spreading very small groups of pixels over several areas. The problem with this method is that since the areas of windthrow were small, getting a pure windthrow pixel was very difficult and most likely the training site pixels that came close to the edge of the windthrow area contained a portion of other surface cover types including the adjacent stand and/or cutblock. Even splitting the framing sites into two windthrow classes didn't alleviate the multi-modal distribution. In order to improve the quality of the windthrow class, larger and more uniform areas of windthrow would be needed. The use of vegetation indices versus using just the six multispectral bands, did improve the supervised classification of windthrow versus non-windthrow in some cases. The best classification was 24ML, which had correctly identified 72.7 % of the high percent windthrow (class 3) plots, 46.7% of the medium percent windthrow (class 2) plots and 61 40% of the low percent windthrow (class 1) plots. The overall image classification accuracy was also quite high at 84.2%. The input channels used in this classification were those of the Model 4 logistic analysis, which was the one with the best diagnostic measures (highest R 2, lowest AIC and -21ogL), though it did not have the highest accuracies when tested on the test data or full data sets. The vegetation indices from the logistic Model 4 included R13, which was based only on the visible wavelength bands (B1-B3), the ratio vegetation index (R4) and R18, both of which incorporate the visible wavelengths and near-infrared (B4), and three Vis which included the short-wave infrared bands (R23, R26, and R37). Four of these Vis were also found to be significant in at least one other logistic model. The Vis making up this model favour the near infrared and short wave infrared regions, which have been associated with changes in leaf water content (Hardisky et al. 1983; Cohen 1991b), and R26 was specifically found to be responsive to changes in plant water stress (Cohen 1991b). It is suspected that the areas of windthrow are differentiated from areas of non-wmdthrow by the changes in leaf water content and an increase in the exposure of bark. Areas of recent windthrow, where new vegetation has not initiated and the foliage of the windthrown trees has dried, would show not only the greatest change in leaf water content compared to adjacent stands or cutblocks but also a high amount of exposed bark, which would explain the higher success rate in detecting these areas. The classification attempts using the Model 4 variables for both classification methods had high success rates for finding the plots with a high percentage of windthrow. This would indicate that using the variables from the best model based on the diagnostic measures does improve the results over using just the bands. Perry and Lautenschlager (1984) suggested that some Vis are functionally equivalent. In their study, two indices were considered equivalent if the decisions based on one index could have been equally well made on the basis of the other index. Though this study used the Landsat MSS bands, which differ in wavelength to the Landsat ETM bands, the principle can still be applied. They found that an index consisting of the ratio of two bands is functionally equivalent to the transform vegetation index using the same bands. 62 ( B5-B4 Thus R26 is functionally equivalent to R37 +0.5 , both of which were V B5+B4 significant in logistic Model 4. Additionally, R37 was the only index that occurred in more than 2 models. This would indicate that Bands 5 (short-wave infrared) and Band 4 (near infrared) could be the most significant bands in detecting windthrow. In order to compare the effects of the site variables, such as vegetation height, canopy closure, and the level of windthrow by volume, to the overall success of the classification attempts, the percentage of classifications that successfully identified each plot was used. The success rate of finding the plots with windthrow depends on a complex interaction between the site and canopy variables, rather than just one. In some cases, even plots with high windthrow levels, and low canopy closure and low vegetation height were only identified by a few of the classification attempts, while other plots with low volumes of windthrow, and high vegetation or canopy closure were identified by more of the classification attempts. This suggests that there are, perhaps, other variables that were not taken into consideration, which were predominant in the training pixels for windthrow, but not present in all other windthrow plots. Such variables could include the amount or type of exposed soil, or the type of vegetation cover in the windthrow area. The high windthrow plots were consistently classified with higher accuracy than the other two classes by the various classification attempts, showing that higher percentages of windthrow are more likely to be correctly identified in a classification than low percentages of windthrow. Additionally, the large plots, which had been placed in larger areas of windthrow were also more consistently classified as windthrow. Ground features similar or smaller in size to the pixel resolution, as the small plots were, are not consistently detectable, with the exception of linear features (Townshend 1981), such as roads, rivers, or long strips of windthrow along the edge of a cutblock. Pixels along complex boundaries and linear features are more likely to be misclassified (Cherrill et al. 1994). As windthrow tends to be a linear feature on the boundary between cutblocks and forest stands, this may serve to explain why some of the small plots were not detectable, despite their having high levels of windthrow and low canopy cover. No relationship was found between the location of the plots on the cutblock edge (i.e. North, South, East, or West) and the detection using the supervised classification, other than 63 that some plots on the southern edge fell into shadow and were not detectable. The predominant factor that seemed to determine the success of the classification was the size and severity of the windthrow area. Small areas were simply not detectable using supervised classification of satellite data with 30m pixel resolution. The high level of structural heterogeneity (i.e., varying canopy heights and presence of snags) of the cedar stands resulted in a high spectral heterogeneity, which also showed a higher level of misclassification as windthrow than the other cover types. This is consistent with what Smith et al. (2002) found in their study, that small patch size results in lower classification accuracy, and higher landscape heterogeneity also leads to lower classification accuracy. Given the accuracies obtained from the self-classification of the framing sites (overall image classification accuracy), the M L classifier was better than the M D classifier. The classification success of the training sites, however, did not relate to the success of the windthrow plot detection. Generally, there was no consistency in the method (MD or ML) that performed the best in the classification of windthrow, as in some cases the M D classifier was best, and in other cases the M L classifier was best. The classifications using the three class dataset failed, most likely due to the few framing pixels for windthrow. According to Hixson et al. (1980), all cover types must be adequately represented by a sufficient number of pixels in order to guarantee a true representation of the spectral characteristics of the classes. This is especially important for the maximum likelihood classifier where the assumption of multivariate normality is made (Schowengerdt 1983). In the case of a classification which uses the logistic Model 4 variables, there would be six input channels, requiring a minimum of seven pixels per training class, but for accurate class parameters to be calculated there should be 60. The two windthrow classes only contained 12 and 14 pixels, which could have contributed to the failure of these classifications. The windthrow areas are, by nature, variable, as the amount of vegetation, exposed wood, and soil, will change from site to site. The windthrow classification of both the two-class and three-class models would likely be improved significantly with the availability of larger windthrow areas that could be used as training sites. 64 5.3 A L T E R N A T I V E A P P R O A C H E S TO D E T E C T I N G W I N D T R H O W Another approach to finding areas of windthrow could be through change detection, or multi-temporal analysis. Multi-temporal analysis has the potential to show where changes on cutblock boundaries occur; however, this requires having two or more images acquired over a period of time, usually at one year increments. Ideally, the images should be acquired at the same time each year, as seasonal differences will affect the success of change detection, as will other changes from one image to another, including the dryness of the soil or vegetation cover. Also, the images need to be perfectly registered, or the results will be impossible to interpret. The disadvantage of this method is that acquiring data of the same area may be difficult or costly, depending on the satellite and weather conditions. Spectral linear unmixing is another technique which has met with some success in other studies. A n attempt at spectral linear unmixing was made using the imagery from this study. The results and discussion of this procedure are presented in Appendix 1. Several issues arose surrounding this approach. Primarily, samples of the pure spectral reflectance of the surface cover types, also known as end-members, must be obtainable. In spectral unmixing, this is even more important than for supervised classification, as the linear unmixing model assumes that each pixel is made up of a linear combination of the end-members. As difficulties arose from finding enough windthrow training sites for supervised classification, it was even more difficult to find pure windthrow end-members for the spectral unmixing. However, as this technique has met with some success in several studies, including the detection of mountain pine beetle attacked trees (Murtha et al. 2000), it would merit further study in its application to windthrow. IKONOS data had been acquired as part of this study, and attempts were made to do a similar analysis as had been done for the Landsat data, however, several issues arose from the use of the IKONOS data. Firstly, the satellite is off-nadir viewing, resulting in image data that is slightly warped from this effect. Despite image correction, it was impossible to accurately register the vector data to the IKONOS image. Attempts were also made to overlay two IKONOS scenes taken three months apart. It was found that 65 overlaying them was impossible, as among other things, several cutblocks were significantly different in size. It was discovered that this was due to the fact that the two scenes were acquired at different view angles, and this effect had not been fully removed through image correction. Additionally, the higher resolution, though an advantage in some instances, proved to make the supervised classification difficult as many more training areas were needed, and more thorough ground-truthing would have been required. IKONOS also lacks the two short-wave infrared bands that Landsat has, limiting the number of vegetation indices and band combinations that could be used. Some of these issues could most likely be resolved with a little work, and it is possible that higher resolution satellite data, though more expensive than Landsat, could prove more effective in the detection of windthrow. These techniques were not attempted as part of this study for several reasons. As only one other Landsat image of the study area without cloud cover was obtained, it was felt that the multi-temporal analysis would have been very limited. For the spectral linear unmixing it was felt that, as it was already difficult to find good training areas for the supervised classification, and linear unmixing is even more sensitive to the quality of the endmembers (training pixels), the results would likely be inaccurate. Attempts had already been made to resolve the issues surrounding the IKONOS data; however, the results of this were unsatisfactory, and likely other image correction processes would need to be used. 66 6 CONCLUSIONS For this study, using a combination of probabilities from logistic regression with supervised classification produced the best results for detecting large areas of high density windthrow. Newer areas and more extensive areas of windthrow were more easily detected than older areas. The logistic regression models using the two-class dataset for detecting windthrow area were only marginally successful. Though the diagnostic measures were fairly low, the results were logical, in that the pixels with higher probabilities of windthrow were concentrated around the cutblock and road edges, and in the more heterogeneous forest types, such as the cedar-dominated (CH) stands, and along the rivers. The three-class models failed to accurately predict windthrow severity. This was most likely due to the lack of training site pixels that could be used to build the models. Windthrow patches in this study area are small, and the few areas large enough to use as training sites could not account for the variability that is inherent in the spectral reflectance of windthrow. In the supervised classification attempts, when selected vegetation indices were combined with logistic probability estimates, the windthrow detection accuracies were higher than for the classifications using just the six multispectral bands. Though the severity of windthrow could not be detected from the Landsat 7 ETM data, using even a combination of the logistic regression models and supervised classification, areas of severe windthrow were detected with accuracies up to 73.6%. This would suggest that Landsat data could be used quite effectively to monitor large (>30m by 30m) areas of windthrow, and to detect new areas of severe windthrow. This would allow forest managers to make decisions relating to salvage or treatment of cutblock edges, monitoring windthrow areas, and improve the windthrow risk modeling. 67 Several aspects of this study suggest further research. The use of vegetation indices should be explored more thoroughly on overall classification accuracies of supervised and unsupervised classifications. With an unbiased method of accuracy assessment for all the cover classes within an image, the results may be in favour of using vegetation indices to improve overall image classification accuracy. Unsupervised classification may also be significantly improved by the inclusion of VTs. Additionally, the use of Vis in other analysis techniques, such as spectral linear unmixing, or the creation of pan-enhanced images could be explored. Multi-temporal analysis of the image data could be attempted, to determine if windthrow can be identified with better success through change detection. The major problem that arose in this study was the lack of windthrow training sites. The use of higher resolution data, such as IKONOS, could improve the results, as the higher pixel resolution would allow smaller areas of windthrow to be used as training sites. With refinement, the techniques and methods explored in this study show potential to be used in monitoring of forest health issues and improving windthrow risk modeling. Furthermore, several aspects of this study merit further investigation as they could improve not only the results of this study, but show potential to improve other techniques in remote sensing. 68 LITERATURE CITED Allen, T. R. and J. A. Kupfer. 2001. Spectral response and spatial pattern of Fraser fir mortality and regeneration, Great Smokey Mountains, USA. Plant Ecology 156(1):59-74. Anderson, G. L., J. D. Hanson, and R. H . Haas. 1993. Evaluating Landsat Thematic Mapper derived vegetation indices for estimating above-ground biomass on semiarid rangelands. Remote Sensing of Environment 45:165-175. Asrar, G., M . Fuchs, and E. T. Kanemasu. 1984. Estimating absorbed photosynthetic radiation and leaf area index from spectral reflectance of wheat. Agronomy Journal 76:300-306. Blackburn, P., J. A . Petty, and K. F. Miller. 1988. A n assessment of the static and dynamic factors involved in windthrow. Forestry 61(l):29-43. Brown, J. L. 1977. Etude de la perturbation des horizons du sol par un arbre qui se renverse et de son impact sur la pedogehese. Canadian Journal of Soil Science 57:173-186. Cherrill, A . J., A. Lane, and R. M . Fuller. 1994. The use of classified Landsat-5 Thematic Mapper imagery in the characterization of landscape composition: a case study in northern England. Journal of Environmental Management 40:357-377. Chomitz, K. M . and D. A . Gray. 1996. Roads, land use, and deforestation: a spatial model approach to Belize. World Bank Economic Review 10:487-512. Cihlar, J., L. St. Laurent, and J. A . Dyer. 1991. Relation between the normalized difference vegetation index and ecological variables. Remote Sensing of Environment 35:279-298. Cohen, W. B. 1991a. Chaparral vegetation reflectance and its potential utility for assessment of fire hazard. Photogrammetric Engineering and Remote Sensing 57(2):203-207. Cohen, W. B. 1991b. Response of vegetation indices to changes in three measures of leaf water stress. Photogrammetric Engineering and Remote Sensing 57(2):195-202. Collins, J. B. and C. E. Woodcock. 1996. A n assessment of several linear change detection techniques for mapping forest mortality using multispectral Landsat T M data. Remote Sensing of Environment 56:66-67. Eid, T. and B.-H. Oyen. 2003. Models for prediction of mortality in even-aged forests. Scandinavian Journal of Forest Research 18:64-77. Elvidge, C. D. and Z. Chen. 1995. Comparison of broad-band and narrow-band red and near-infrared vegetation indices. Remote Sensing of Environment 54:38-48. 69 Fiorella, M . and W. J. Ripple. 1993. Determining successional stage of temperate coniferous forests with Landsat satellite data. Photogrammetric Engineering and Remote Sensing 59(2):239-246. Foody, G. M . and R. A . Hil l . 1996. Classification of tropical forest classes from Landsat T M data. International Journal of Remote Sensing 17(12):2353-2367. Galinski, W. 1989. A windthrow-risk estimation for coniferous forests. Forestry 62(2):139-146. Gunter, J. T., D. G. Hodges, C. M . Swain, and J. L. Regens. 2000. Predicting the urbanization of pine and mixed forests in Saint Tammany Parish. Photogrammetric Engineering and Remote Sensing 66(12):1469-1476. Hardisky, M . A., V. Klemas, and R. M . Smart. 1983. The influence of soil salinity, growth form, and leaf moisture on the spectral radiance of Spartina alterniflora canopies. Photogrammetric Engineering and Remote Sensing 49(l):77-83. Helliwell, D. 1989. Tree roots and the stability of trees. Arboricultural Journal 13(3):243-248. Hixson, M . M . , N . Scholz, T. Fuhs, and T. Akiyamal. 1980. Evaluation of several schemes for classification of remotely sensed data. Photogrammetric Engineering and Remote Sensing 46(12):1547-1553. Hosmer, D. and S. Lemeshow. 2000. Applied logistic regression. 2nd ed. John Wiley and Sons, Inc., New York. Hudak, J. 1993. Detection and classification of forest damage using remote sensing. In Proceedings of the International Forum on Airborne Multispectral Scanning for Forestry and Mapping, Apri l 13-16,1992, Val-Morin, Quebec, p. Huete, A. R. 1989. A soil-adjusted vegetation index (SAVT). Remote Sensing of Environment 25:295-309. Jackson, R. G , G. M . Foody, and C. P. Quine. 2000. Characterizing windthrown gaps from fine spatial resolution remotely sensed data. Forest Ecology and Management 135:253-260. Jordon, C. F. 1969. Derivation of leaf area index from quality of light on the forest floor. Ecology 50:663-666. Kahabka, H. , M . Dees, B. Koch, and N . Saidani. 2001a. Subject: Sturm Lothar: Schadenserfassung mit optischen Fernerkundungsdaten aus forstlicher Sicht. [in German with English abstract.] http://www.dkkv.org/forum2001 /Datei58.pdf. Accessed July, 16, 2003. Kahabka, H. , G. Ramminger, N . Saidani, M . Dees, and B. Koch. 2001b. Schadenserfassun nach Orkan "lothar" mit methoden der Fernerkundung. [In German with English abstract.]. AFZ-Der Wald 25: 1331-1333 70 Klinka, K., J. Pojar, and D. V. Meidinger. 1991. Revision of biogeoclimatic units of coastal British Columbia. Northwest Science 65(l):32-47. Knick, S. T., J. Rotenberry, and T. Zarriello. 1997. Supervised classification of Landsat Thematic Mapper imagery in a semi-arid rangeland by nonparametric discriminant analysis. Photogrammetric Engineering and Remote Sensing 63(l):79-86. Koutsias, N . and M . Karteris. 1998. Logistic regression modeling of multitemporal Thematic Mapper data for burned area mapping. International Journal of Remote Sensing 19(18):3499-3514. Krajina, V. 1965. Biogeoclimatic zones in British Columbia. Ecology of Western North America 1:1-17. Lewis, T. 1982. Ecosystems of the Port McNeill Block (Block 4) of the tree farm licence 25. Western Forest Products Ltd., Port McNeill, B.C. Lewis, T. 1985. Ecosystems of the Quatsino tree-farm Licence (TFL) 6. Western Forest Products Ltd., Vancouver, B.C. Lillesand, T. M . and R. W. Kiefer. 2000. Remote sensing and image interpretation. 4th ed. John Wiley and Sons, Inc., New. York. Lloyd, D. 1990. A phenological classification of terrestrial vegetation cover using short-wave vegetation index imagery. International Journal of Remote Sensing 11:2269-2279. Ludeke, A . K , R. C. Maggio, and L. M . Reid. 1990. A n analysis of anthropogenic deforestation using logistic regression and GIS. Journal of Environmental Management 31:247-259. McDonald, A. J., F. M . Gemmell, and P. E. Lewis. 1998. Investigation of the utility of spectral vegetation indices for determining information on coniferous forests. Remote Sensing of Environment 66:250-272. Mertens, B. and E. F. Lambin. 2000. Land-cover-change trajectories in Southern Cameroon. Annals of the Association of American Geographers 90(3):467-494. Mitchell, S. J. 1995. A synopsis of windthrow in BC. Occurrence, implications and management. Pages 448-459 in M . P. Courts and J. Grace, eds, Wind and Trees. Cambridge University Press, New York. Mitchell, S. J., T. Hailemariam, and Y. Kulis. 2000. Empirical modeling of cutblock edge windthrow risk on Vancouver Island, Canada, using stand level information. Forest Ecology and Management 154(1-2):117-130. Mollard, J. and J. R. Janes. 1984. Airphoto interpretation and the Canadian landscape. Canadian Government Publishing Centre, Quebec. 71 Muchoney, D. M . and A. H . Strahler. 2002. Pixel- and site-based calibration and validation methods for evaluating supervised classification of remotely sensed data. Remote Sensing of Environment 81:290-299. Murtha, P. A . 2000a. Monitoring cutblocks, riparian strips and windthrow on northern Vancouver Island with RADARS AT F2 data. In B.C. RADARSAT Workshop, May 9, 2000, Victoria, B.C., p. http: / /207.162.296.217/ publications /pdf/rsat Zbc214.pdf. Murtha, P. A. 2000b. Monitoring riparian leave strips with multi-temporal RADARSAT C-band satellite data. In Proceedings of a Conference on the Biology and Management of Species and Habitats at Risk, Feb. 15-19,1999, Kamloops, B.C., p. 10. Murtha, P. A . 2000c. Surficial geology and climatic effects on forest clearcut tone in RADARSAT images of northern Vancouver Island. Canadian Journal of Remote Sensing 26(3):142-151. Murtha, P. A. 2001. Radarsat Monitoring windthrow decimation of riparian strips, Northern Vancouver Island. In Windthrow Assessment and Management in British Columbia, Windthrow Workshop, January 31 to February 1, Richmond, B.C., p. 111-121. Murtha, P. A., Z. Bortolot, and J. Thurston. 2000. A Landsat T M spectral unmixing mountain pine beetle attack-fraction map in the Vanderhoof Forest District, British Columbia. In 22nd. Annual Canadian Remote Sensing Symposium, August 21-25, University of Victoria, p. 6. Nelson, B. W., V. Kapos, J. B. Adams, W. J. Oliveira, O. P. G. Braun, and I. L. do Amara. 1994. Forest disturbance by large blowdowns in the Brazilian Amazon. Ecology 75(3):853-858. PCI. 2002. Geomatica. 8.2.1 ed. PCI, Richmond Hil l , Ontario. Pearson, R. L. and L. D. Miller. 1972. Remote mapping of standing crop biomass for estimation of productivity of the Shortgrass Prairie. In Proceedings of the 8th International Symposium on Remote Sensing of Environment, Oct. 2-6, Ann Harbour, Michigan, p. 1457-1481. Peltonen, M . 1999. Windthrow and dead-standing trees as bark beetle breeding material at forest-clearcut edge. Scandinavian Journal of Forest Research 14:505-511. Perestrello de Vasconcelos, M . J., S. Silva, M . Tome, M . Alvin, and J. M . C. Pereira. 2001. Spatial prediction of fire ignition probabilities: comparing logistic regression and neural networks. Photogrammetric Engineering and Remote Sensing 67(1):73-81. Perry, C. R. and L. F. Lautenschlager. 1984. Functional equivalence of spectral vegetation indices. Remote Sensing of Environment 14:169-182. 72 Prescott, C. E., ed. 1996. Salal cedar hemlock integrated research program. Research update #1: December 1996. Faculty of Forestry, University of British Columbia, Vancouver, B.C. Purevdroj, T., R. Tateishi, T. Ishiyama, and Y. Honda. 1998. Relationship between percent vegetation cover and vegetation indices. International Journal of Remote Sensing 19(18):3519-3536. Ramminger, G., M . Dees, and B. Koch. 2001. Subject: STURMMON Schadenserfassung mit Radardaten aus forstlicher Sicht - Erste Ergebnisse. [in German with English abstract.] http://www.dkkv.org/forum2001 /Datei63.pdf. Accessed July 16, 2003. Richardson, A. J. and C. L. Wiegand. 1977. Distinguishing vegetation from soil background information. Photogrammetric Engineering and Remote Sensing 43:1541-1552. Richardson, W., A . Pentland, R. Crane, and H . Horwitz. 1976. Number of signatures necessary for accurate classification. In Symposium on Machine Processing of Remotely Sensed Data, June 29 - Julyl, 1976, Purdue University, Indiana, p. 3A.28-23A.34. Rouse, J. W., R. H . Hass, J. A . Schnell, and D. W. Deering. 1973. Monitoring vegetation systems in the great plains with ERTS, N A S A SP 351. In Proceedings of the 3rd Symposium on ERTS-1, Dec. 10-14,1973, Wahington, D . C , p. 309-317. Ruel, J. 1995. Understanding windthrow: silvicultural implications. The Forestry Chronicle 71(4):434-445. Sader, S. A . 1995. Spatial characteristics of forest clearing and vegetation regrowth as detected by Landsat Thematic Mapper imagery. Photogrammetric Engineering and Remote Sensing 61(9):1145-1151. SAS Institute Inc. 1999. SAS OnlineDoc. Vol. 2003. Version 8 ed. SAS Institute Inc., Cary, NC. Schowengerdt, R. A . 1983. Techniques for image processing and classification in remote sensing. Academic Press, Inc., Orlando, Florida. Sieben, B. 2001. Mapping and Investigating the mean and extreme wind regime of coastal British Columbia. In Windthrow Assessment and Management in British Columbia, Proceedings of the Windthrow Workshop, January 31 to February 1, Richmond, B.C., p. 91-96. Smith, J. H. , J. D. Wickham, S. V. Stehman, and L. Yang. 2002. Impacts of patch size and land-cover heterogeneity on thematic image classification accuracy. Photogrammetric Engineering and Remote Sensing 68(l):65-70. Snee, R. S. 1977. Validation of regression models: methods and examples. Technometrics 19(4):415-428. 73 Stathers, R. J., T. P. Rollerson, and S. J. Mitchell. 1994. Windthrow handbook for British Columbia forests. British Columbia Ministry of Forests, Victoria, B.C. Story, M . H . and J. B. Campbell. 1986. The effect of training data variability on classification accuracy. In ACSM-ASPRS Annual Convention. Technical Papers, March, Washington, D . C , p. 370-379. Swain, P. H . and S. M . Davis, eds. 1978. Remote Sensing - The Quantitative Approach. McGraw-Hill, New York. Townshend, J. G. 1981. Effects of spatial resolution on the classification of land cover type. In Ecological Mapping from Ground, Air and Space. Institute of Terrestrial Ecology Symposium No. 10, Nov. 25-27, Monks Wood Experimental Station, Great Britain, p. 101-112. Trigg, S. and S. Flasse. 2002. A n evaluation of different bi-spectral spaces for discriminating burned shrub-savannah. International Journal of Remote Sensing 22(13):2641-2647. Watts, S. B., ed. 1983. Forestry Handbook for British Columbia. 4th ed. Forestry Undergraduate Society, University of British Columbia, Vancouver, British Columbia. Weetman, G. F., R. Fournier, E. Schorbus Panozzo, and J. Barker. 1990. Post-burn nitrogen and phosphorus availability of deep humus soils in coastal British Columbia cedar/hemlock forests and the use of fertilization and salal eradication to restore productivity. In Sustained Productivity of Forest Soils, Proceedings of the 7th North American Forest Soils Conference, July, 1988, Vancouver, B.C., p. 525. 74 APPENDIX 75 A P P E N D I X I R E P O R T O N S P E C T R A L L I N E A R U N M I X I N G Spectral Linear Unmixing Analysis of the Study Area Pixels along the edge of a cutblock are frequently spectrally mixed with signatures from more than one object. Spectral mixing occurs when objects (eg trees, water) with different spectral properties are represented by a single pixel. This problem can be alleviated by using a technique called spectral mixture analysis or linear spectral unmixing. Spectral unmixing is a mathematical approach to determuriing the percentage of pure elements for each object (trees, water) called endmembers that are found within the pixel. Mathematically, the idea behind linear spectral unmixing can be written as: 1=1 Where D N b is the D N value of the pixel being examined in band b, F ; is the fraction of endmember i , D N i , b is the D N value of endmember i in band b, and £b is the error associated with band b. The object of unmixing is to predict F i for each endmember over the entire image using least-squares estimates. Spectral unmixing should allow for the detection of sub-pixel patches of blowdown. The following endmembers were selected from the same training areas as were used for the supervised classification: • Windthrow • New cut • H B • CW • Shadow 76 These endmembers were selected because the windthrow pixels would most likely contain a mixture of these elements as they are commonly found adjacent to windthrow. The spectral signature of these endmembers rwas calculated from the pixels selected to represent the cover types. The output of the subpixel analysis was a pseudo-coloured image for each endmember, where the black areas of the image represent the pixels containing a low percentage of that end-member, and the red areas represent pixels that contain a high percentage of that endmember. The results of the spectral unmixing show a high correlation between the old-growth cedar stand types and the windthrow endmember. This is not an unexpected result, as the cedar stand types display a fairly open canopy structure, and the presence of "candelabra" cedars, which have dead tops. The exposed wood from these cedars would likely have a very similar spectral reflectance to that of windthrow. The assumption with Subpixel analysis is that the spectral signatures for each endmember represent the pure spectra of that cover type. As the spectra for the endmembers were calculated from pixels selected from the image, there will likely be some contamination to the endmembers. For the four non-windthrow cover types this was not so much an issue, as large areas were available for careful pixel selection and it was felt that good endmembers had been obtained. However, the endmember for windthrow was likely not as good for two reasons. The first reason was that large areas of windthrow were very rare, and only one area (8 pixels large) was felt to be "pure" enough to serve as an end-member. The second reason is that windthrow is a fairly variable event, as was discovered during the field work phases. Each area of windthrow was composed of varying amounts of exposed soil, bark and vegetation, depending on the initial stand density, original under story vegetation, and age of windthrow. This, essentially, means that there can be no single endmember to represent windthrow, rather endmembers representing exposed soil, and bark would be necessary. Pixels with a certain ratio of exposed soil and bark could then be identified as windthrow. Though it may be possible (with further fieldwork) to find pixels representing exposed soil, it would likely be impossible to find an area as large a 30 by 30 metres containing only exposed bark. The IKONOS image cannot be used to obtain the endmember for bark, as 77 it does not have the two short-wave infrared bands, and the wavelengths of the visible and near-infrared bands is slightly different from that of Landsat. Thus, the only way to obtain the endmember for bark would be to use a spectrophotometer in the field, to gather that information. The results could be improved with the availability of better endmembers for the windthrow class, and also with hyper-spectral imagery. Several studies have had success with using Landsat imagery in conjunction with spectral unmixing, however, this technique is usually applied to hyper-spectral imagery as this allows more endmembers to be used. 78 Appendix 1 Windthrow endmember image of study area, where red pixels have a high percentage of windthrow in the pixel, and black pixels have a low percentage of windthrow. 7 9 

Cite

Citation Scheme:

        

Citations by CSL (citeproc-js)

Usage Statistics

Share

Embed

Customize your widget with the following options, then copy and paste the code below into the HTML of your page to embed this item in your website.
                        
                            <div id="ubcOpenCollectionsWidgetDisplay">
                            <script id="ubcOpenCollectionsWidget"
                            src="{[{embed.src}]}"
                            data-item="{[{embed.item}]}"
                            data-collection="{[{embed.collection}]}"
                            data-metadata="{[{embed.showMetadata}]}"
                            data-width="{[{embed.width}]}"
                            async >
                            </script>
                            </div>
                        
                    
IIIF logo Our image viewer uses the IIIF 2.0 standard. To load this item in other compatible viewers, use this url:
http://iiif.library.ubc.ca/presentation/dsp.831.1-0075082/manifest

Comment

Related Items