Open Collections

UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Evaluation of Landsat thematic mapper data for reforestation assessment Bansal, Arun Kumar 1988

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

Item Metadata


831-UBC_1988_A6_7 B36.pdf [ 3.67MB ]
JSON: 831-1.0075291.json
JSON-LD: 831-1.0075291-ld.json
RDF/XML (Pretty): 831-1.0075291-rdf.xml
RDF/JSON: 831-1.0075291-rdf.json
Turtle: 831-1.0075291-turtle.txt
N-Triples: 831-1.0075291-rdf-ntriples.txt
Original Record: 831-1.0075291-source.json
Full Text

Full Text

EVALUATION OF LANDSAT THEMATIC MAPPER DATA FOR REFORESTATION ASSESSMENT BY ARUN KUMAR BANSAL M.Sc.(Physics), University of Rajasthan, 1974 Member of Indian Forest Service, 1975 A THESIS SUBMITTED IN PARTIAL FULFILMENT OF THE REQUIREMENTS FOR THE DEGREE OF MASTER OF SCIENCE in THE FACULTY OF GRADUATE STUDIES (Department of Forestry/Remote Sensing) We accept this thesis as conforming to the required standard THE UNIVERSITY OF BRITISH COLUMBIA OCTOBER, 1988 © Arun Kumar Bansal, 1988 In presenting this thesis in partial fulfilment 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. ARUN KUMAR BANSAL Department of Forest Resource Management The University of British Columbia 2357 Main Mall Vancouver, Canada V6T 1W5 Date 31 October, 1988. ABSTRACT Forests are important natural resources of Canada. Their renewal has been recognized to be important for continued wood supply and for other benefits. Consequently, the major emphasis of forest management activities focuses upon restocking clearcut forest lands. Effective planning and successful implementation of reforestation programs require efficient techniques for obtaining timely and accurate information regarding restocking status over cutover forest lands. In this thesis the potential of Landsat thematic mapper (TM) data for monitoring reforesting clearcuts was investigated. Landsat-5 TM data covering clearcut forest lands reforesting with lodgepole pine (Pinus contorta Dougl.) were analyzed. To assess spectral separability of various restocking classes and classifying reforestation areas according to their stocking status multivariate distance measures were employed to select the optimum three band subset from six reflective TM bands. Three commonly used vegetation indices, namely the ratio vegetation index, the normalized difference vegetation index, and the infrared index, were also studied for quantitative assessment of vegetation. The main conclusion of the study is that TM bands 3, 4, and 5 are the best for discriminating various restocking classes. The classification accuracy was estimated to be approximately 90 percent. The infrared index appears to be the most suitable vegetation index for quantitative assessment of reforestation. ii TABLE OF CONTENTS I ABSTRACT ii II TABLE OF CONTENTS iii LU LIST OF TABLES v IY LIST OF FIGURES vi V LIST OF ABBREVIATIONS vii VI ACKNOWLEDGMENTS viii 1. INTRODUCTION 1 2. HTERATURE REVIEW 4 2.1 BASIC PRLNCLPLES 4 2.1.1 Characteristics of thematic mapper bands 4 2.1.2 Spectral properties of vegetation 9 2.2 REFORESTATION MONITORING USING LANDSAT DATA 11 2.2.1 Delineation and mapping of clearcuts 11 2.2.2 Assessment of vegetation amount 14 2.3 DIGITAL IMAGE ANALYSIS 18 2.4 RESEARCH OBJECTIVES 24 3. MATERIALS AND METHODS 25 3.1 STUDY SITE 25 3.2 GROUND DATA 25 3.3 SATELLITE DATA 27 3.3.1 Image description 27 3.3.2 Image analysis 28 3.3.3 Quantitative assessment of reforestation 30 iii 4. RESULTS AND DISCUSSION 32 4.1 ESTIMATION OF GROUND COVER PERCENT 32 4.2 ANALYSIS OF THEMATIC MAPPER DATA 41 4.2.1 Band correlations and data dimensionality 41 4.2.2 Mean digital number curves and feature space plots 45 4.2.3 Optimum band selection 48 4.2.4 Classification accuracy 54 4.2.5 Vegetation indices 55 5. CONCLUSIONS 62 IJTERATURE CITED 65 iv LIST OF TABLES I Spectral ranges of TM bands 6 II Principal applications of TM bands 8 III Commonly used vegetation indices 16 IV ANOVA for simple linear regression between GCP and PS 34 V ANOVA for conditional linear regression between GCP and PS 34 VI ANOVA for the Lack of Fit test 37 VLT Forest stockings of the study areas 39 Vm Assignment of study areas to GCP classes 40 LX Correlation matrix of the reflective TM bands 42 X Eigenvector matrix of the sub scene TM data 43 XI Optimum Index Factor values of all possible three band subsets of the six TM bands 52 Xn Pairwise B-distances for the GCP classes 53 Xm Confusion matrix based on cluster groups of the reforestation areas 56 XTV Radiometric calibration coefficients for TM data 57 XV Coefficients of determination between vegetation indices and GCP 58 v LIST OF FIGURES 1 Regions of electromagnetic spectrum used in passive remote sensing 5 2 Generalized spectral reflectance curves for green leaf and soil 10 3 Percent correct classification versus B-distance 23 4 Location map of the study site 26 5 Scatter diagram of the observations 33 6 Residuals from conditional linear regression 36 7 Regression line and 95 percent confidence bands 38 8 Graph of the eigenvalues 44 9 Mean digital number curves of various GCP classes 46 10 Coincident spectral plots 47 11 Feature space plot of GCP classes (bands 5 and 4) 49 12 Feature space plot of GCP classes (bands 5 and 7) 50 13 Feature space plot of GCP classes (bands 5 and 3) 51 14 Infrared Index curve 59 15 Normalized Difference Vegetation Index curve 60 16 Ratio Vegetation Index curve 61 vi LIST OF ABBREVIATIONS ASCII American Standard Code for Information Interchange ANOVA Analysis of Variance CCRS Canada Centre for Remote Sensing DN Digital Number ERTS Earth Resource Technology Satellite FIRMS Forest Information for Resource Management Systems GCP Ground Cover Percent GVI Green Vegetation Index IRI Infrared Index LAI Leaf Area Index LCV Laboratory for Computational Vision LOF Lack of Fit MDN Mean Digital Number MOSAICS Multi-Observational Satellite Image Correction System MSS Multi-Spectral Scanner NDVI Normalized Difference Vegetation Index NIR Near Infrared OIF Optimum Index Factor PS Photo Stocking PVT. Perpendicular Vegetation Index SWIR Short Wave Infrared TIE Thermal Infrared TM Thematic Mapper TMS Thematic Mapper Simulator TVI Transformed Vegetation Index UBC University of British Columbia vii ACKNOWLEDGMENTS I would like to express my sincere gratitude to my supervisor, Dr. Peter A Murtha, and the members of my committee, Dr. Karel Klinka, and Dr. Jack Thirgood, for their support and for a few enlightening discussions which have guided my work. I wish to thank Mr. Ken Day, Resident Forester, Alex Fraser Research Forest, for his assistance in the conduct of field work. I also wish to acknowledge the able assistance of the technical staff at the Faculty of Forestry, namely Raoul J. Wiart and Ms. Nedenia M. Krajci, and Tim Lee, Programmer at the Laboratory for Computational Vision. I am grateful to the Canadian Commonwealth Scholarship and Fellowship Committee for the financial support that made my stay in Canada possible, and to the Government of India, Ministry of Environment and Forests, for granting me study leave. The funding for this research was provided from a block grant from the Canadian Forestry Service (now administered by the Natural Science and Engineering Research Council of Canada) to Faculty of Forestry, University of British Columbia. viii 1 CHAPTER 1 INTRODUCTION The evolution of spaceborne remote sensing has been a major technological advance that has opened new avenues for the acquisition of information about the earth's resources. Launch of the Earth Resource Technology Satellite, ERTS-1, on July 23, 1972 marked the beginning of the application of space technology for obtaining earth's resource information. The ERTS-1, later renamed as Landsat-1, was the first satellite of its kind specifically designed for natural resource monitoring. Subsequently, satellite data have been increasingly utilized to obtain vital information about the earth's land and water resources. Remote sensing techniques have found world wide application in forest management. These techniques offer unique advantages over conventional aerial and ground survey methods. Satellite techniques furnish synoptic repetitive coverage in the visible, photographic or near infrared (NTR), short wave infrared (SWTR) and thermal infrared (TTR) regions of the electromagnetic spectrum. Availability of satellite data in digital format makes them suitable for computer assisted analysis thereby enhancing information extraction. The Multi-Spectral Scanner (MSS) was the primary sensor system on board Landsats-1, -2 and -3. MSS provided spectral reflectance data at 79 meter ground resolution in four spectral bands covering the visual and NIR regions. These data were widely used for broad forest cover type mapping and monitoring. However, spatial and spectral resolution limitations of Landsat MSS data tended to preclude any detailed stocking assessment over reforesting 2 forest clearcuts. Vegetation monitoring was an important factor in the specification of the Thematic Mapper (TM) sensor system which became operational with the launch of Landsat-4 in July, 1982 and remained the most important sensor system on Landsat-5. TM offers significant improvements over its MSS forerunner in the areas of spatial, spectral and radiometric resolutions. Preliminary studies have revealed the potential of TM data for extending the usefulness of remote sensing techniques in detailed assessment of the forest resources (Crist and Cione, 1984; Ahern and Archibald, 1986; Horler and Ahern, 1986). Monitoring of new growth over clearcut lands is of great importance for efficient forest planning and forest resources development of any nation. It is of even greater significance in Canada where forestry is passing through a transition phase, from exploitation to systematic management. Consequently, one of Canada's major concerns for forest management focuses upon restocking cutover forest lands. Assessment of forest stocking through conventional aerial/ground survey methods is very time consuming and expensive. A recent analysis entitled "The backlog of unstocked forest land in British Columbia and the impact of reforestation programs" emphasizes the inadequacy of the information about the extent of accumulated backlog of non-satisfactorily restocked areas (Pearse et al. 1986a). In another companion study Pearse et al. (1986b) concluded: Indeed our investigation of the reforestation issue in British Columbia has left us with the impression that the information and data gathering systems are not well attuned to the major policy questions at stake. The most basic question facing policy-makers is whether we are planting enough trees to meet the objectives; but statistical information does not lend itself to an answer to this question nor does it throw much light on whether the seedlings are being planted in the right place, whether they are surviving, or whether lands left unplanted will restock naturally and other questions basic to effective assessment of policies and programs. 3 Remotely sensed data acquired through Landsat's TM scanner seem to have the potential for providing answers to some of these questions, particularly the ones relating to the extent and location of non-satisfactorily restocked forest lands and the performance of planted and naturally reforesting clearcuts. The research described in this thesis is an evaluation of Landsat TM data for the classification of clearcut forest lands according to forest stocking. A methodology, based upon standard techniques for analysis of remotely sensed digital data, is proposed for selecting the optimum band combination for maximum discrimination among various restocking stages. Vegetation indices suitable for quantitative assessment of forest stocking have also been identified. To proceed with the presentation of this research, Chapter 2 provides the reader a basic understanding of TM sensor system, and fundamental principles of remote sensing technology and digital image analysis. In Chapter 2 the objectives of the research are also stated. The techniques for ground truth data collection and analysis of TM data are discussed in Chapter 3. Study results and a discussion are presented in Chapter 4. Conclusions are summarized in chapter 5, including concerns about future research. 4 CHAPTER 2 LITERATURE REVIEW 2.1 BASIC PRINCIPLES Remote sensing of the earth's resources involves the detection and measurement of electromagnetic radiations that are reflected or emitted by various features on the earth's surface. The wavelength regions of electromagnetic spectrum which are used in passive remote sensing1 of the earth's resources are shown in Figure 1. Spectral reflectance curves2, also known as spectral signatures, of different object types are the primary keys to a features detection and identification. In remote sensing spectral signature of any object is defined as relative spectral distribution of the reflected or emitted radiation (Sievers and Kriebel, 1980). The purpose of this section is to briefly discuss the characteristics of TM bands and the spectral reflectance properties of vegetation. 2.1.1 Characteristics of thematic mapper bands TM is designed to acquire integrated reflectance data about the earth's surface in seven narrow spectral bands. Wavelength regions of TM bands are shown in Table I. 1 In passive remote sensing naturally available energy is sensed by the detectors in contrast to active sensor systems, such as RADAR, which have their own energy source to illuminate features of interest. 2 Curves between spectral reflectance and the wavelengths. Spectral reflectance is the ratio of the radiant energy reflected by a body to that incident upon it at specified wavelength interval, represented in percent. Figure 1. Regions of electromagnetic spectrum used in passive remote sensing CO z o 5 LU-CE < CC o Ul 0. CO (0 S Ul <fl > CO CC o z Ul CO Ul -J o > < Ul m w > z a < £ 5 < O Ul £ z £ CC £ — < Ul J -I K Q K 2 2 T < u. T T 10 M U L T I S P E C T R A L S C A N N E R S I I I I I III 100 1000 /urn W A V E L E N G T H * — • NOTE: Adapted from Lillesand and Kiefer, 1987. * The wavelength scale is logarithmic. Table I. Spectral ranges of TM bands Band Number Wavelength Electromagnetic Designation (in uma) Region 1 0.45 • • 0.52 Blue 2 0.52 • 0.60 Green 3 0.63 • • 0.69 Red 4 0.76 • 0.90 NTR 5 1.55 • 1.75 SWLRI 6 10.40 • - 12.50 TTRb 7 2.08 • • 2.35 SWLR n SOURCE: Landsat Data Users Notes, NASA (1982) a micrometer (10"6 meters). b The TTR band sensors record emitted radiations. The other six bands, collectively known as reflective TM bands, have sensors which record reflected radiations. 7 TM images the earth's surface with 30 meter x 30 meter ground resolution, known as pixel3, except band 6 which has a ground resolution of 120 meter x 120 meter. The sensors record reflected (or emitted in case of band 6) electromagnetic radiation within their designated wavelength bands as 256 intensity levels (0 being no reflectance, 255 being the maximum). For each pixel, TM data consist of seven digital numbers (DNs), one for each band, representing integrated radiant energy reflected (emitted for band 6) from the earth's surface. TM sensor system is described in detail in Freden and Gordon (1983). The reflective TM bands (bands 1, 2, 3, 4, 5 and 7) have been selected to optimize their information content for assessing the condition of vegetation. Bands 1 and 3 utilize the radiances from vegetation which are largely determined by leaf pigment concentrations. Band 3 is also an excellent in vivo chlorophyll band (Tucker, 1978a). Band 2 is placed to maximize spectral information content about the green reflectance peak of living vegetation. Band 4 senses in the region of the NTH plateau and is considered to be good for general vegetation applications including detection of vegetation growing under certain types of stresses which cause the rounding of the NIR reflectance plateau (Tucker, 1978b). Bands 5 and 7 record in the region of absorption by leaf water content of green vegetation. Band 5 is responsive to both water content and cellular structure of the leaves and is therefore more suitable for vegetation applications, while band 7, characterized by intense water absorption, is suited for geological remote sensing (Elvidge and Lyon, 1985). Band 6 is designed for sensing emitted radiations for thermal mapping. Some principal applications of TM bands are shown in Table LL 3 The term pixel is derived from picture element denoting spatial resolution of the sensor system. 8 Table LI. Principal applications of TM bands Band Number Applications Designation 1 coastal water mapping, differentiation of soil from vegetation and coniferous from deciduous vegetation 2 measurement of green reflectance peak of vegetation for vigour assessment 3 chlorophyll absorption band for plant species discrimination 4 biomass determination dnd water body delineation 5 vegetation moisture measurement, differentiation of snow from clouds 6 thermal mapping, plant heat stress analysis 7 discrimination of rock types, hydrothermal mapping SOURCE: Freden and Gordon, (1983). 9 2.1.2 Spectral properties of vegetation Reflected energy from the earth's surface results from the interaction of incident solar radiations with the objects on the ground. The interactions of solar radiation and green leaves in different regions of the electromagnetic spectrum are well documented: spectral properties of plants (Gates et al., 1965); physical and physiological properties of plants (Gates, 1970); reflectance and transmittance of light by leaves (Wolley, 1971); leaf reflectance of near infrared (Gausman, 1974); reflectance of leaf components (Gausman, 1977); distinguishing succulent plants from crop and woody plants (Gausman, et al., 1978). The generalized spectral reflectance curves of green leaves and soil are shown in Figure 2. For the spectral reflectance curve of vegetation, low values in the blue (TM band 1) and the red (TM band 3) regions represent strong electronic absorption by leaf chlorophyll and other pigments. Middle level reflectance in the green (TM band 2) region is due to low absorption of green light by chlorophyll. High reflectance in the NTH (TM band 4) region results from Fresnel reflections4 from the leaf cellular structure (cell wall - air space interfaces) and cellular constituents (Gausman, 1974). Two distinct regions of low reflectance centred at 1.4 um5 and 1.9 um in the SWTR (TM bands 5 and 7) region are due to strong vibrational absorption by leaf water content. Total leaf spectral reflectance depends upon a number of factors including plant species, age of the foliage, time in growing season, moisture content and nutrient status of the cells. Spectral reflectance properties of vegetation canopies in a field situation are modified, both quantitatively and qualitatively, 4 Reflections at the interfaces of refractive index discontinuities. 5 micrometer equal to 10"6 meters. 10 Figure 2. Generalized spectral reflectance curves for green leaf and soil NOTE: Adapted from Hoffer (1978). 11 from those of leaves due to variations in leaf orientations, shadows, and background surfaces, such as soil. Knipling (1970), Colwell (1974), Jarvis et al. (1976), and Curran and Milton (1983) have discussed these effects in detail. In the case of reforesting lands, which are partially covered by young trees, a TM pixel is an assemblage of different components, mcluding vegetation canopies comprising mainly of leaves along with other plant structures, background soil, and shadows in various proportions. As the vegetation grows, the spectral reflectance gradually changes from predominantly that of the background soil to that of the vegetation. The spectral reflectance curves for soil and vegetation (Figure 2) suggest a pattern of these changes. In the visible and SWIR regions soils are more reflective than green leaves and spectral reflectance should decrease with an increasing proportion of vegetation cover. Changes would be reverse in the NIR region due to higher leaf reflectance. In addition, coniferous vegetation also results in increasing amounts of shadows and old needles which modify these changes. Generally, an increasing proportion of shadows and old needles reduce the overall canopy reflectance at all wavelengths. 2.2 REFORESTATION MONITORING USING LANDSAT DATA Since the availability of the first satellite data from Landsat-1, reforestation monitoring has been an important research topic. Study areas can be grouped into two main categories. First, the delineation and mapping of forest clearcuts, and second, the assessment of vegetation amount. 2.2.1 Delineation and mapping of clearcuts Investigations based upon interpretation of black and white or color photographic products generated from Landsat MSS digital data indicated the 12 possibility of monitoring the progression of forest depletions (Zsilinszky, 1973; Oswald, 1974; Lee, 1975; Murtha and Watson, 1975; Bansal, 1986). Aldrich (1975) highlighted the importance of the season for data acquisition as being a critical factor for separating coniferous and deciduous vegetation. The development of digital image analysis techniques has enhanced information interpretation from digital data. Supervised classification of multi-temporal6 MSS band 5 (0.6 um to 0.7 um) overlays were used to consistently detect 2 hectares or larger clearcut areas (Derenyi et al., 1984; Rencz, 1985). Performing supervised classification on single date imagery, Bryant et al. (1979) identified clearcuts down to 3 hectares. Reasonable areal assessments of forest clearcuts were achieved using Landsat MSS data. Crapper (1980) formulated that the errors in area assessment were a function of the size and shape of the individual blocks. He estimated the relative error to be of the order of 2 percent for areas of approximately 50 hectares. This increased to roughly 10 percent for individual areas of approximately 2 hectares. Archibald and Ahern (1985) performed a statistical analysis of forest clearcut mapping accuracy of visual delineation of depletion boundaries on two date composite images. Their analysis revealed that the root mean square boundary placement errors were slightly more than one half of the ground resolution of the sensor system. The depletion areas were found to be over-estimated by 2 to 10 percent, higher estimation errors being associated with smaller individual areas. Spectral reflectance of clearcuts was found to show variability related to the age of regrowth (Bryant et al., 1979; Shimabukuro et al., 1980). However, classification of clearcuts into detailed resource categories using MSS data 6 Multi-temporal, also known as multi-date, refers to two or more images of the same area obtained by the same sensor system on different dates. 13 were not very reliable (Johnson and Barthmaier, 1979). TM data with improved spatial, spectral and radiometric resolutions were expected to provide more detailed information for delineating new conifer growth. In an analysis of the quality and information content of Landsat-4 TM and MSS data, Anuta et al. (1984) concluded that principal component analysis7 indicated four significant dimensions in TM data compared to two dimensions of the MSS data. Further, the spectral analysis based upon averaged transformed divergence revealed twice as many separable classes with TM compared to MSS data. Williams et al. (1984), and DeGloria (1984) also found forest information content of TM data to be higher than that of MSS data. TM data were found to improve detection, delineation and mapping of small forest clearings due to a lower proportion of boundary pixels and sharper boundary definitions among different cover types (DeGloria, 1984; Irons et al., 1985). Hopkins (1988) reported average forest/non-forest classification accuracies to be about 98 percent. Additional spectral information provided by TM data, particularly in bands 5 and 7, seems to be important for detailed assessment of forest stocking status in early regeneration stages. Studying the forest depletion and regeneration information content of TM data covering Dryden-Lac-Seul region in North-Western Ontario, Horler and Ahern (1986) observed: The visible reflectance shows a decrease with time, as slash and soil in the clearcut are progressively covered with green vegetation. The effect is more pronounced in TM 1 and 3, which is understandable given the greater chlorophyll absorption in these bands. TM 4 shows an increase in reflectance, again understandable as a consequence of green vegetation. The reflectance decreases with time in the SWTR (TM 5 and 7). 7 Principal component analysis is also referred to as the Hotelling or Karhunen-Loeve transformation (Singh and Harisson, 1985). 14 However the size of decrease relative to the variability is greatest in this region. This may indicate that the SWIR is a particularly good region for monitoring regeneration. They concluded that TM SWIR bands may perform better, both in discriminating new clearcuts from surrounding forest and in separating clearcuts of different ages. Ahern and Archibald (1986) found that a range of forest conditions, including identification of clearcuts and regeneration, was visible by combining TM bands 3, 4, and 5. They concluded that high SWTR reflectance was consistently associated with a low proportion of living ground cover. Werle et al. (1986) analyzed Landsat TM data for forest depletion monitoring on Vancouver Island. Applying a Z test on single band mean digital numbers (MDNs) as a measure of separability, they observed that most recent clearcuts were distinct on bands 3 and 5 while 15 to 20 year old regeneration sites were distinct on band 4. Evidence would suggest that Landsat TM data can be used for delineation and mapping of forest clearcuts and for obtaining accurate estimates of forest depletions. Most of the studies dealing with classification of clearcuts have investigated their separability based upon the age of the clearcuts. However, from the resource management perspective, discrimination of clearcuts based upon actual restocking status would be more meaningful. 2.2.2 Assessment of vegetation amount Quantitative assessment of vegetation based upon its spectral reflectance properties was first reported by Jordon (1969) who used NLR/Red ratio values to derive leaf area index (LAI)8 for forest canopies in a tropical rain forest. Since Jordan's (1969) study, many workers have observed LAI is the one-sided area of leaves per unit area of ground. 15 relationships between remotely sensed visible and NIR radiance and vegetation amount, resulting from characteristically high reflectance differences between regions of absorption and reflectance. The majority of vegetation assessment studies have been carried out on agricultural, grassland or rangeland environments. Vegetation amount is represented by plant biomass, LAI or green LAI. Although deterministic (Chance, 1981) or stochastic (Goel and Thomson, 1984) models would be ideal to describe the relationships between radiance data and the parameters representing vegetation amount, the presence of many unknown and immeasurable variables only allows for the formulation of empirical models. Experimental models, usually regression-based relationships, have been found valid for a wide range of scenes and sensors (Tucker, 1979; Ajai et al., 1983; Ihse and Graneli, 1985). A number of spectral indices, known as vegetation indices, most of which are either ratios or linear combinations of red and NIR radiances, have been developed for crop monitoring using Landsat MSS data. Some of the commonly used vegetation indices are given in Table ILL Normalized Difference Vegetation Index (NDVE), also known as vegetation index (Rouse et al., 1973; Tucker et al., 1979; Curran, 1983), is the most popular index due to its simplicity and high degree of standardization, and the fact that it makes no assumptions regarding the statistical distribution of the data (Curran, 1980). Furthermore, from the theoretical standpoint, by modelling the canopy reflectance using the radiative transfer equations, NDVE has been shown to be positively correlated to the vegetation amount computed as LAI and ground cover fraction (Choudhury, 1987). There have been very few studies using satellite data in assessing restocking over clearcuts mainly due to the fact that the Landsat MSS data were not found suitable for detailed discrimination of revegetation stages. Table in. Commonly used vegetation indices 16 Name Formula References RVP NLR / Red Jordon, 1969 NDVP TVIC GVId PVIe (NLR - Red) / (NLR + Red) (NDVI + 0.5)m - 0.29 MSS4 - 0.56 MSS5 + 0.60 MSS6 + 0.49 MSS7 [(RecU - Red™^ + (NLRMi] - NIRveg)2] 1/2 Pearson and Miller, 1972 Rouse et al., 1973 Kauth and Thomas, 1976 Richardson and Wiegand, 1977 a Ratio Vegetation Index b Normalized Difference Vegetation Index c Transformed Vegetation Index d Green Vegetation Index 8 Perpendicular Vegetation Index 17 Townshend et al. (1983) reported that TM bands 5 and 7 seem to contain different information about vegetation compared to that contained by the red and NLR bands of the MSS. Horler and Ahern (1986) also found that the SWER bands provide significant additional information about forests in early stages of regeneration. Running et al. (1986) found a high correlation between LAI of coniferous forests and NTR/Red radiance ratio obtained from airborne TM data. They observed that this ratio can provide a methodology for quantitative study of vegetation disturbances. Investigating the correlation of percent canopy closure with the response of individual thematic mapper simulator (TMS)9 bands for selected forest sites in the San Juan National Forest, Butera (1986) determined TMS bands 1, 5, and 7, the bands not covered by Landsat MSS, to be the most important. Butera considered percent forest canopy closure to be an appropriate vegetation assessment parameter due to the fact that it is relevant to wildlife habitat assessment, watershed run off estimation, and other forest management activities. At the ground resolution of TM data, percent vegetation cover has been considered the most important vegetation variable for studying reflectance characteristics of sparsely forested areas (Graetz and Gentle, 1982; Wilson and Tueller, 1987). This is due to the fact that composite reflectance (p) of an area comprising of vegetation and soil background can be expressed as: Ay PV + Ag pg p = (i) A 9 TMS is an airborne scanner, maintained by the National Aeronautics and Space Administration's National Space Technology Laboratory, designed to simulate the Landsat TM. 18 where, A v = vegetation area, Ag = soil area, pv = vegetation reflectance, ps = soil reflectance, and A = Ay 4" Ag (Janza, 1975). Studying the microclimatic alterations produced by forest fires, L6pez-Garcia et al. (1986) determined that monitoring index10 was more suitable than NDVI to monitor the reforestation of fire affected areas. A similar index, termed as infrared index11 (IRI) by Gross et al. (1986), was studied by Curran and Williamson (1987). The IRI showed slightly higher correlation with green LAI compared to the NDVI. The foregoing discussion establishes that TM data contain an order of magnitude more information than do the Landsat MSS data in regard to forest regrowth, particularly in the early stages of revegetation. This thesis continues in the direction of utilizing the additional information contained by Landsat TM data for more efficient planning and management of the forest resources, in particular, clearcut areas. 2.3 DIGITAL IMAGE ANALYSIS TM data consist of a set of seven DNs for each pixel. DNs for each pixel constitute a measurement vector, also known as spectral pattern, which defines a point in 7-dimensional measurement space, called feature space. Digital classification of multi-band image data is performed using pattern recognition 1 0 (NIR - SWIR II) / (NIR + SWIR II) 1 1 (NIR - SWIR I) / (NIR + SWIR I) 19 techniques. In these techniques the feature space is partitioned into decision regions based upon the information contained in the spectral patterns of the classes of interest. Of the many mathematical methods used in pattern-recognition, the statistical methods are particularly appropriate in remote sensing applications (Swain, 1978). These methods make use of the probability density functions associated with different pattern classes. Probability density functions are usually unknown and are extracted from a set of training patterns12. In most remote sensing applications probability functions can be approximated by multivariate normal density functions. The classifiers designed on this assumption are found to be robust in the sense that classification accuracies are not very sensitive to even moderately severe violations of this assumption (Swain, 1978). Accordingly, any pattern class is completely defined by its mean vector and covariance matrix. Classification is then accomplished by assigning the pixels to the pattern classes following statistical decision rules such as the maximum likelihood or the nearest neighbor decision rules. Probability density functions of the classes of interest may actually overlap due to within class variability resulting from inherent randomness of nature. Consequently, classification has an associated probability of error depending upon the degree of overlap. Furthermore, there is a degree of redundancy and high cost in terms of the speed and complexity of classification due to interband correlations exhibited by multi-spectral data. Although classification accuracy is generally reduced when fewer than all the available bands are used, it is usually possible to select a subset of all bands which optimize a trade off between classification accuracy and computational speed The spectral patterns of pixels representing the classes of interest. 20 and cost. A study conducted by Fu et al. (1970) shows that subsets may be almost as effective as the complete feature set. In another study by Fu et al. (1969), multi-class agricultural crop classification experiments showed that subsets of TM bands result in better classification performance than that achieved by the complete set of bands. This was attributed to deviation of actual probability density functions from the assumed normal distributions. Choosing a maximum of three bands is most logical since not more than three bands can be viewed at any one time. A widely followed approach for feature selection13 depends on the concept of statistical distance measures between the probability density functions characterizing the pattern classes. Divergence14, transformed divergence15 and B-distance18 (also known as Jeffreys-Matusita distance) are some of the commonly used measures of separability between a pair of classes. Comparing the various statistical separability measures as feature selection criteria Swain and King (1973), and Swain (1978) found B-distance to be better than divergence or transformed divergence. 1 3 Feature selection refers to finding an optimum subset of multi-band data. 1 4 Divergence is closely related to Shanon's logarithmic measure of information. Divergence as applied in remote sensing is discussed in Kailath (1967), and Fu et al. (1970). 1 5 Transformed divergence is defined as: TD = 2[1 - eD / 8 ], where, D is the divergence. (Swain et al, 1971). 1 6 B-distance is related to Bhattacharyya coefficient (Bhattacharyya, 1942) which is a measure of affinity between two statistical populations. Some properties of B-distance are discussed in Kailath (1967). 21 B-distance (By) for a pair of Gaussian classes17 is given by: B,- = [2(l-e- a)] (ii) where, a = + v 2 ioge [(i K2, + Zj)/2} i)/(i ay i i cy i n {uj} = mean vector for class i, {Ujj = mean vector for class j, {2y = covariance matrix for class i, {£} = covariance matrix for class j, and I (A) I = determinant of matrix {A}. The first term in the expression of a is akin to the multivariate form of square of the normalized distance between the mean vectors. The second term represents the contribution due to the differences in the covariance structures of the two classes which is equal to zero only when the two covariance matrices are equal, a increases monotonically with the difference between the class probability density functions. Negative exponential in equation (ii) results in exponentially decreasing weights to increasing differences between the class probability density functions. Consequently, B-distance has a saturating behaviour like the probability of correct classification, saturation value being 2172. For two classes with equal a priori probabilities and B-distance B, error probability, Pe, based upon maximum likelihood classifier, is given by the following expression (Kailath, 1967): (Swain, 1978). (1 - (B2 (1 - V4 B2))m) < 2P e < (1-V 2B 2) (iii) Classes having normal probability density functions. 22 Upper and lower bounds of percent correct classification (Pe = 100(l-Pe)) as a function of B-distance are shown in Figure 3. The combination giving largest B-distance is an obvious choice if there are only two classes of interest. In a multi-class situation with m classes, there will be m(m - l)/2 pairs of classes. The commonly used decision rule in such cases is to maximize the average separability (Min et al., 1968; Fu, 1971; Swain and King, 1973; Card and Angelici, 1983). Average B-distance (Bavg) for a m-class situation can be calculated as: 2 B a v g = 2;Z, By (iv) m (m - 1) (i = 1 to m - 1; j = i + 1 to m), where, Bu is the B-distance between classes i and j. An alternative strategy is to maximize the minimum separability, also known as maximin criterion (Grettenberg, 1963; Fu and Chen, 1965; Swain, 1972; Card and Angelici, 1983). This criterion aims at selecting the combination that maximizes distance between the hardest-to-separate or the least separable class pair. However, since the relationship between B-distance and classification accuracy is nonlinear, widely separable classes contribute too much to B a v s when compared with the less separable classes. Consequently, the combination selected on the basis of maximum Ba V B may not be the best. Conversely, maximin procedure selects the combination which is best for separating the hardest-to-separate class pair but there is no guarantee that this combination would provide optimum classification results (Kumar, 1977). For selecting the best three band subset from the six reflective TM bands it would be necessary to compute B-distances for twenty possible three band combinations. The number of candidate combinations can be greatly 23 Figure 3. Percent correct classification versus B-distance 24 reduced by ranking them according to their information content. Chavez et al. (1982) developed the Optimum Index Factor (OIF) technique for ranking all possible three band combinations. OIF for a three band combination is computed as follows: OIF = (v) Zj ICQ I (i = 1 to 3; J = 1 to 3), Where, SDt = standard deviation for band i, and I CCj I = absolute value of the correlation coefficient between any two of the three bands. OIF weighs the variances of individual bands by using their standard deviations and correlation between them determined by the correlation coefficients. Since the variance for a band is a measure of its information content, the combination having the largest OIF value should contain the most information with the least amount of duplication. Since there is often little difference between closely ranked combinations, within two or three positions (Chavez, 1984), B-distances of the first few combinations can be compared for making the final choice of the optimum three band subset. 2.4 RESEARCH OBJECTIVES This study is an evaluation of Landsat TM data for detailed assessment of reforesting clearcuts. The main goal was to determine the information content of TM data for reforestation monitoring. Specific research hypotheses are : (i) Landsat thematic mapper data can be used to discriminate different stages of restocking over reforesting clearcuts, and (ii) Vegetation indices are suitable parameters for quantitative assessment of stocking over reforesting clearcuts. 25 CHAPTER 3 MATERIALS AND METHODS 3.1 STUDY SITE The study site is located near the Gavin Lake block of Alex Fraser Research Forest of the University of British Columbia (UBC), approximately 52°26'N to 52°36'N latitudes and 121°46'E to 121°54'E longitudes (Figure 4). Eleven clearcut parcels reforesting with lodgepole pine (Pinus contorta Dougl.), and four areas of mature lodgepole pine forests were selected for this study. All the areas are situated on virtually flat terrain, maximum slope 10 percent, and are classified as Sub-boreal spruce (SBSkx) subzone according to the biogeoclimatic ecosystem classification followed in British Columbia (Anon., 1987a; Pojar et al., 1987). The reforesting areas provided a range of restocking stages, from very recent clearcuts to areas with 15 year old regeneration. 3.2 GROUND DATA The study areas were grouped into restocking classes based on ground reference information (ground truth) of forest stocking as obtained from regression relationships between ground measured ground cover percent (GCP) and photo stocking (PS) interpreted from medium scale (1:10,000) normal-color aerial photographs. GCP denotes the proportion of ground area covered by planted and/or natural lodgepole pine seedlings represented as percent. PS refers to forest stocking as interpreted from aerial photographs. The study site aerial photographs, taken on August 27, 1986, were acquired from the British Columbia Ministry of Forests. Study areas were outlined on the aerial photographs and were divided into 3 mm x 3 mm grids. 26 Figure 4. Location map of the study site 27 The grid cells, hereafter referred to as photo plots, represented 30 meter x 30 meter squares on ground, corresponding to the spatial resolution of TM data. PSs of all the photo plots were interpreted using a 2X-pocket stereoscope. All the photo plots were grouped into eleven PS classes. Fifty photo plots, representing all PS classes uniformly, were selected randomly for ground measurements. Photo plots were located on the ground using a radial triangulation method. Average crown diameters, at the point of maximum spread, of all conifer seedlings in each plot were measured. GCP was calculated presuming circular crowns. No distinction was made between lodgepole pine and other conifer species. A relationship between PS and GCP was derived through regression analysis. Bansal and Murtha (1988) have shown that regression relationship between GCP and PS can be used for obtaining reliable estimates of ground reference data for analysis of TM data. GCPs for individual areas were computed as the regression estimates for the average PSs of the respective areas. 3.3 SATELLITE DATA 3.3.1 Image description A Landsat-5 TM seven band quarter-scene was acquired by the Forest Information for Resource Management Systems (FIRMS, the Remote Sensing Laboratory in the UBC Faculty of Forestry) for forest management related studies. The scene was imaged on August 17, 1986 at 1022 hours local time (path 47, row 24, quadrant 1; scene identification number 50899182203) and was cloud free. Since radiometric quality of the image data received originally 28 in 1986 was uncertain, system corrected geocoded18 computer compatible tapes reproduced on MOSAICS (Multi-Observational Satellite Image Correction System)19 were acquired from the Canada Centre for Remote Sensing (CCRS) in May 1988. 512 pixel x 512 pixel sub scene image files, one each for the six reflective TM bands20, covering the study area, were extracted using the image analysis facilities at the Laboratory for Computational Vision (LCV) in the UBC Department of Computer Science. The image files were subsequently transferred to FIRMS. 3.3.2 Image analysis Analyses of the image data were carried out at FLRMS using a microcomputer based image analysis system MERLDLAN/PC21 and other data analysis packages. Correlations between TM bands were studied by computing a band correlation matrix for the sub scene. The program PRLNCOM, developed by Dumoulin (1985) was used for principal component analysis of the six band TM images to determine the intrinsic dimensionality of the data. The number 1 8 The system corrected geocoded imagery was produced with along and across scan systematic geometric corrections (without ground control points), and systematic (or relative) radiometric corrections. 1 9 MOSAICS is a precision image correction facility at the Canada Centre for Remote Sensing for processing satellite data. A brief description of this image correction system is available in Friedel and Fisher (1987) and MOSAICS Technical Information Package (Anon., 1987b). 2 0 The TLR band data was not analyzed in this study. 2 1 MERLDLAN/PC is a microcomputer based image analysis system developed by MacDonald, Dettwiler and Associates, Richmond, British Columbia, Canada V6X 2Z9. 29 of useful dimensions were determined through the scree-test22. Study areas were outlined on the sub scene image and map overlay files (masks) were generated. Pixel location (X and Y coordinates) and DNs (one for each of the six TM bands) were extracted for each study area as ASCII (American Standard Code for Information Interchange) files using in-house programs. MDNs and covariance matrices were computed for each GCP class. MDNs for individual study areas were also calculated. A preliminary analysis of TM bands suitable for discriminating various GCP classes was performed by plotting MDN curves23 and feature space plots24. Feature space plots were drawn as bivariate ellipses, using SYGRAPH (Wilkinson, 1988a) for each restocking class. Ellipses for each class are centred at MDNs with major axes equal to the standard deviations for any two bands. The orientation of the ellipses are determined by the covariance between the two bands. Differences in MDNs and degree of overlap between the ellipses for different classes provided initial indications regarding separability of the GCP classes and suitability of different bands in discriminating various GCP classes. Final selection of the best three band subset was accomplished through statistical separability analysis using B-distance (section 2.3) as a measure of 2 2 According to the scree-test, proposed by Cattell (1966), the number of eigenvectors to be extracted is determined to be not more than the point at which the eigenvalue graph begins to level off forming a straight line with almost horizontal slope (Dillon and Goldstein, 1984). 2 3 MDN curves are plots of MDNs of various restocking classes against corresponding mean band wavelengths. The MDN curves can be termed as inferred spectral response curves, where spectral responses are inferred from the corresponding MDNs. 2 4 Feature space plots are scatterplots of responses of various classes in the measurement space. 30 separability between a pair of classes. The data were checked for deviations from the normal assumption by plotting single band histograms for all GCP classes. All possible three band combinations were ranked in descending order of their information content employing the OIF technique (section 2.3). Highest ranking combinations with similar OLF values were selected for computation of B-distance. B-distance for all possible pairs of classes were calculated for all the selected combinations. Both the average distance and maximin criteria were investigated to choose the optimum band combination. Assessment regarding discrimination of GCP classes, using the selected three bands, was done by calculating achievable classification accuracies using equation (iii). Suitability of the selected bands for classifying the areas into various restocking classes was further evaluated by performing independent cluster analysis on the spectral data of these bands. ' SYSTATs K-means clustering (Wilkinson, 1988b) was used to assign each pixel of the study areas to one of four groups. K-means clustering splits a set of records into a selected number of groups in order to maximize between - relative to within - group variation. Cluster groups were reassembled maintaining spatial locations, as obtained from pixel extraction, to produce study area diagrams. The cluster classes were compared pixel by pixel with the assumed correct classification (classes based on average GCPs) by analyzing the confusion matrix, also known as confusion table (Singh, 1980). 3.3.3 Quantitative assessment of reforestation Suitability of TM data for quantitative assessment of stocking over reforesting clearcuts was investigated by performing correlation analyses between average GCPs of the study areas, estimated in section 3.2, and vegetation indices computed from TM data. The following vegetation indices 31 were studied: (i) Ratio Vegetation Index TM 4 RVI = . (vi) TM 3 (ii) Normalized Difference Vegetation Index TM 4 - TM 3 NDVI = . (vii) TM 4 + TM 3 (iii) Infrared Index TM 4 - TM 5 IRI = : (viii) TM 4 + TM 5 where, TM 3 , TM 4, and TM 5 are at satellite spectral radiances (L) in the respective bands calculated from the corresponding MDNs using the following equation: I^  = a MDNi - bt (ix) where, § = gain (w m 2 ster1 um'1 per DN), = bias (w m 2 ster1 um1), and i = TM band number. (Tassan, 1987). Since previous studies by Colwell (1974), Tucker (1979), and Wardley and Curran (1984) reveal that vegetation indices have nonlinear relationship with vegetation parameters, correlation between the vegetation indices and GCP was analyzed using SYSTATs nonlinear estimation module, NONLIN (Wilkinson, 1988b). Correlation coefficients and rates of change of vegetation indices (slopes of the best fit curves) were compared to determine the most appropriate vegetation index for monitoring of regrowth over clearcut forest lands. 32 CHAPTER 4 RESULTS AND DISCUSSION 4.1 ESTIMATION OF GROUND COVER PERCENT Scatter diagram of the measured GCPs and PSs, as interpreted from aerial photographs of the fifty photo plots (Figure 5), suggested a linear relationship between GCP and PS. ChiSquare test for goodness of fit (Snedecor and Cochran, 1980) and Bartlett's test for heterogeneity of variances (Pearson and Hartley, 1966) showed that the observations satisfy the basic assumptions of regression that residuals are normally and independently distributed with zero mean and a common variance. Simple linear regression using least square technique gave following significant relationship (analysis of variance, ANOVA, is given in Table TV): GCP = - 0.84475 + 1.032877 * PS (R-SQUARE 2 5 = 0.9889) (x) The negative intercept in this relationship indicates that when PS is zero, estimated value of GCP will be less than zero. However, in actual situations, the minimum possible GCP is zero. A two tailed t-test resulted in the acceptance of the null hypothesis intercept equal to zero at 95 percent confidence level. Hence, it was decided to derive a conditional relationship with the restriction intercept equal to zero. Linear regression through the origin resulted in the following significant relationship (ANOVA is given in Table V): 2 5 R-SQUARE is the coefficient of determination, and represents the proportion of variation in the response variable explained by the model. Figure 5. Scatter diagram of the observations 34 Table IV. ANOVA for simple linear regression between GCP and PS Source of Degrees of Sum of Mean Variance Variation Freedom Squares Squares Ratio (F) Regression 1 23681.61 23681.61 4278.20* Residual 48 265.70 5.54 Total 49 23947.31 * Significant at 0.01 level of significance. Table V. ANOVA for conditional linear regression between GCP and PS Source of Degrees of Sum of Mean Variance Variation Freedom Squares Squares Ratio (F) Regression 1 54345.78 54345.78 9488.12' Residual 49 280.66 5.73 Total 50 54626.44 * Significant at 0.01 level of significance. 35 GCP = 1.013094 * PS (R-SQUARE = 0.9883 26) (xi) Appropriateness of the linear model was verified by plotting the residuals, and performing a statistical lack of fit (LOF) test. Plot of residuals against estimated GCPs (Figure 6) did not show any discernable pattern. LOF test was accomplished by partitioning the residual variance into sum of squares due to pure error and LOF (Weisberg, 1985). Pure error was computed as pooled variance multiplied by the corresponding degrees of freedom. Comparison of the calculated F Ratio (Table VI) with tabulated F value (2.82, for 10 and 39 degrees of freedom at 0.01 level of significance) showed that LOF error was not significantly greater than error due to random variability of the observations. These tests confirmed the correctness of the linear model represented by equation (xi). The regression line and 95 percent confidence bands around it are graphically represented in Figure 7. Since the regression line is forced to pass through the origin, there is no estimation error for PS equal to zero. Maximum error was approximately ± 2 GCP units. Equation (xi), accordingly, provided very precise estimates of GCP. Average GCPs of the study areas are given in Table VLT. The study areas were grouped into five classes on the basis of average GCPs (Table VLTJ). 2 6 R-Square is based upon the corrected total and regression sum of squares rather than the uncorrected sums of squares shown in the ANOVA table (Table V) so that it is comparable to R-SQUARE for the unconditional regression (Casella, 1983; Myers, 1986). 36 Figure 6. Residuals from the conditional linear regression 0 20 40 60 80 GCP ESTIMATE Table VI. ANOVA for the Lack of Fit test 37 Source of Variation Degrees of Freedom Sum of Squares Mean Squares Variance Ratio (F) Residual 49 280.66 LOF 10 117.29 11.73 2.80* PE 39 163.37 4.19 * Not significant at 0.01 level of significance. Figure 7. Regression line and 95 percent confidence bands Table VII. Forest stockings of the study areas Area No. Year of Year of Photo Ground Logging* Plantation* Stocking Cover (percent) Percent 1 1976 natural6 15.29 15.49 2 1976-77 1984 10.26 10.39 3 1976-77 1984 11.88 12.04 4 1976-77 1984 17.99 18.23 5 1976-77 1984 11.06 11.20 6 1976 1983 11.60 11.76 7 1979,82 naturalb 0.00 0.00 8 1971 1976 42.59 43.14 9 1972-73 1976 46.46 47.07 10 1985 1986 0.00 0.00 11 1985 0.00 0.00 12 c 100.00 13 c 100.00 14 c 100.00 15 e 100.00 a The year of logging and plantation have been taken from the records of the Regional Forest Office, Cariboo Forest Region, Williams Lake, B.C., Canada. b Naturally reforesting areas. c These are mature lodgepole pine forests (age approximately 100 years). No ground measurements were done in these areas and since they appeared to be closed canopy forests on the aerial photographs, their GCPs were taken to be nominally equal to 100 percent. Table VIII. Assignment of study areas to GCP classes GCP class GCP Study areas 1 0 7, 10, 11 2 10-12 2, 3, 5, 6 3 15-18 1, 4 4 43-47 8, 9 5 100 12, 13, 14, 15 41 4.2 ANALYSIS OF THEMATIC MAPPER DATA 4.2.1 Band correlations and data dimensionality Band correlation matrix for the sub scene image (Table LX) showed high correlations (R-SQUARES greater than 0.80) among the three visible bands, TM bands 1, 2, and 3. The two SWIR bands, TM bands 5 and 7, also had high correlation (R-SQUARE equal to 0.83). The NIR band, TM band 4, was least correlated (R-SQUARE less than 0.50) with the other bands. Visible, NIR and SWIR regions seem to be three basically independent spectral regions for remote sensing of vegetation, as was previously observed by Vickers and Brown (1986), and Horler and Ahem (1986). The eigenvector matrix of the standardized principal component transform, based on the correlation matrix, of the sub scene image is shown in Table X. Standardized principal component analysis was performed because previous studies by Singh and Harisson (1985), and Fung and LeDrew (1987) comparing standardized versus non-standardized principal components derived from Landsat MSS data had concluded that the standardized components offer a better description of the underlying data dimensionality. Scree-test (Figure 8) indicated the number of useful eigenvectors to be three. The first three components together contain most the variance (97.50 percent) of the original six band data. Therefore, intrinsic dimensionality of TM data was determined to be three. The first component was a measure of brightness as all the eigenvector coefficients were positive. The second component was essentially a visible to NIR contrast, with the two SWIR bands virtually balancing each other and was responsive to the combination of high NIR reflectance and low visible reflectance, characteristic of green vegetation, and was thus a measure of greenness. The third component was a contrast between the two SWIR and 42 Table LX. Correlation matrix of the reflective TM bands TM band 1 2 3 4 5 7 1 1.0000 2 0.9207 1.0000 3 0.9314 0.9554 1.0000 4 0.5361 0.6579 0.5441 1.0000 5 0.8550 0.9197 0.9017 0.7635 1.0000 7 0.8357 0.8719 0.8896 0.5883 0.9110 43 Table X. Eigenvector matrix of the sub scene data TM Eigenvector (Principal Component) Coefficients Band PCI PC2 PC3 PC4 PC5 PC6 1 0.4134 -0.2980 -0.4841 0.7060 0.0718 -0.0484 2 0.4319 -0.0966 -0.2770 -0.4765 -0.4424 -0.5520 3 0.4254 -0.2923 -0.1203 -0.4126 0.0799 0.7366 4 0.3221 0.8822 -0.1526 0.1127 -0.1808 0.2219 5 0.4320 0.1411 0.2322 -0.1360 0.7877 -0.3171 7 0.4138 -0.1350 0.7728 0.2704 -0.3736 0.0256 Eigenvector Eigenvalue Percent Cumulative Number Variance Percent Variance 1 5.0681 84.47 84.47 2 0.5987 9.98 94.45 3 0.1830 3.05 97.50 4 0.0753 1.25 98.75 5 0.0455 0.76 99.51 6 0.0294 0.49 100.00 44 45 the other four bands and has been termed swirness (Horler and Ahem, 1986) or wetness (Crist and Cione, 1984). The above analysis furnish further evidence in support of a fundamental TM data structure suggested by Crist and Cione (1984) based on the Tasselled Cap transformation (Kauth and Thomas, 1976). 4.2.2 Mean digital number curves and feature space plots The MDN curves (Figure 9) for various stocking classes displayed a discernible pattern. It can be inferred that the spectral radiance in the visible and SWIR regions (TM bands 1, 2, 3, 5, and 7) decreased with rising GCP values. This decrease is attributable to higher absorption of visible and SWIR radiations with increasing GCP. Increasing amounts of shadows and proportions of old needles would also reduce the reflectance in these regions. The magnitude of decrease was larger in bands 5 and 7 than that in bands 1, 2 and 3. The spectral radiance in the NLR region (TM band 4) increased initially, as the vegetation began to grow (GCP going up from 0 to 10), and then very gradually decreased with further increases in GCP. This apparently anomalous behaviour of NLR reflectance was due to the fact that higher NLR reflectance coming from vegetation is quickly taken over by increasing contribution from the shadows which have very low reflectance compared to the green vegetation and background. However, the responses of various GCP classes in different spectral bands overlapped to varying degrees as seen from the coincident spectral plots26 (Figure 10). The size of differences in MDN relative to variability was 2 6 Coincident spectral plots are plots of the mean spectral responses of each GCP class represented by the corresponding MDNs and one standard deviation spreads. Figure 9. Mean digital number curves of various GCP classes 0.4 0.8 1.2 1.6 2.0 2.4 WAVELENGTH (micrometers) NOTE: MDNs for each GCP class are plotted against mid wavelength of the corresponding TM bands. Figure 10. Coincident spectral plots 160 D I G I T A L 120 80-N U M B E 40 R S x X TM 1 GCP CLASSES x x X x x X TM 2 TM 3 X X X • • • x x x x X TM 4 X X TM 5 x X X . TM 7 TM BANDS • M E A N V A L U E * 1 S T A N D A R D D E V I A T I O N 48 highest in band 5 which indicated that TM band 5 was the most important single band in discriminating restocking stages. Overlapping responses of the GCP classes were better revealed by feature space plots (Figure 11 - bands 5 and 4; Figure 12 - bands 5 and 7; Figure 13 - bands 5 and 3). Restocking stages are better distinguished in the feature space defined by TM bands 4 and 5. GCP classes 1, 2 and 3, 4, and 5 were found to be spectrally distinct. Classes 2 and 3, having the maximum overlap, were least separated classes. 4.2.3 Optimum band selection Table XI ranks all possible three band subsets of the six TM bands on the basis of the OIF values. Interestingly, four combinations comprising of bands 4 and 5 and one of the remaining four bands formed a distinct group with high OLF values. This emphasized that bands 4 and 5 were important for discriminating various stocking classes, as was apparent from the feature space plots (section 4.2.2). The four top ranking combinations, in descending order of OLF ranks, were: bands 4, 5 and 7; bands 3, 4 and 5; bands 1, 4 and 5; bands 2, 4 and 5. These combinations were investigated for final selection based on B-distances. Single band histograms confirmed the validity of the multivariate normal approximation for the probability density functions of various GCP classes for computation of B-distances using equation (ii). B-distances for all possible GCP class pairs are summarized in Table XII. Highest average B-distance was 1.273378 for the combinations of bands 3, 4, and 5. The maximin criterion showed the combination of bands 2, 4, and 5 to be the best for discriminating the least separated pair of classes (GCP classes 2 and 3). Figure 11. Feature space plot of GCP classes (bands 5 and 4) NOTE: Ellipses represent one standard deviation(s) of the DNs of the corresponding GCP classes. Figure 12. Feature space plot of GCP classes (bands 5 and 7) BAND 5 NOTE: Ellipses represent one standard deviation(s) of the DNs of the corresponding GCP classes. Figure 13. Feature space plot of GCP classes (bands 5 and 3) BAND 5 NOTE: Ellipses represent one standard deviation(s) of the DNs of the corresponding GCP classes. 52 Table XI. Optimum Index Factor values of all possible three band subsets of the six TM bands BANDS S S A ZCorry OIF 4 5 7 52.5964 2.2630 23.2419 3 4 5 49.3269 2.2093 22.3269 14 5 47.8264 2.1546 22.1973 2 4 5 46.8774 2.3411 20.0237 3 4 7 34.7702 2.0220 17.1959 14 7. 33.2697 1.9601 16.9735 2 4 7 32.3207 2.1181 15.2593 13 4 30.0002 2.0116 14.9136 3 5 7 40.1403 2.7023 14.8541 15 7 38.6398 2.6017 14.8518 2 5 7 37.6908 2.7026 13.9461 2 3 4 29.0512 2.1574 13.4658 13 5 35.3703 2.6881 13.1581 12 4 27.5507 2.1147 13.0282 2 3 5 34.4213 2.7768 12.3960 12 5 32.8941 2.6954 12.2038 13 7 20.8136 2.6567 7.8344 2 3 7 19.8646 2.7169 7.3115 12 7 18.3641 2.6283 6.9871 12 3 15.0946 2.8075 5.3765 Table XR. Pairwise B-distances for the GCP classes 53 GCP Band Combination Classes 4 5 7 3 4 5 14 5 2 4 5 1,2 1.155699 1.154958 1.149142 1.138526 1,3 1.266158 1.271103 1.267737 1.263419 1, 4 1.405902 1.402963 1.404086 1.402512 1,5 1.414212 1.414212 1.414212 1.414212 2, 3B 0.682603 0.695210 0.685283 0.698263 2,4 1.350772 1.346635 1.347793 1.346915 2,5 1.414213 1.414213 1.414213 1.414213 3,4 1.210433 1.206120 1.210011 1.207215 3, 5 1.414213 1.414213 1.414213 1.414213 4,5 1.414197 1.414190 1.414183 1.414189 AVERAGE 1.272840 1.273378 1.272087 1.271368 NOTE: B-distances are calculated for all possible pairs of GCP classes using three of the six reflective TM bands at a time. a least separated class pair. 54 Combination of bands 3, 4, and 5 was selected because the higher wavelength radiations of band 3 are less effected by atmospheric scattering compared to the lower wavelength radiations of band 2. Average B-distance using the six band TM data was equal to 1.30, only marginally higher than the average B-distance for the optimum three band subset, suggesting that the three band subset, comprising of bands 3, 4, and 5, would perform almost as well as the entire six band data in (iiscriminating the various reforestation areas. This combination makes use of spectral responses of vegetation in all three distinct spectral regions: the visible region characterized by high absorption due to leaf pigments, the near infrared region with high reflectance from the leaf cellular structure, and the short wave infrared region where reflectance is largely dominated by absorption due to leaf moisture content. 4.2.4 Classification accuracy B-distances between various restocking classes were employed to compute classification accuracies. Average classification accuracy was estimated to be greater than 90 percent. However, classification errors up to 40 percent were expected in case of the least separated classes, GCP class 2 and 3. Even by using all six band data classification error for these classes was expected to be approaching 35 percent (B-distance, based on six band data = 0.7637). It was clear that discrimination of reforesting areas with very close percent vegetation covers, GCPs within 5, was not feasible. Consequently, for cluster analysis, classes 2 and 3 were merged and GCP classes were redesignated as 1, 2, 3, and 4. Cluster analysis was performed to split the study area pixels into four groups. Study area diagrams showed fairly homogeneous clusters which were 55 closely related to the GCP classes. Heterogeneity in the cluster groups was seen mostly on the boundaries. Maximum variation was seen in recent clearcuts where many pixels were grouped into the class representing 10 to 15 percent vegetation cover. The confusion matrix (Table XLTJ) showed overall agreement to be about 90 percent. ChiSquare test for independence showed that cluster groups were highly dependent upon the assumed correct classification. The clustering technique may, therefore, be an effective method for classifying reforestation areas into several stocking classes. 4.2.5 Vegetation indices Gain and bias coefficients used for converting MDNs of the study areas to at satellite radiances (sections 2.2.2 and 3.3.3) are reproduced in Table XTV. Best fit curves for the three vegetation indices, investigated for finding correlation with vegetation amount, are shown in Figures 14 to 16. All the three indices, LRI, NDVI, and RVI, were highly correlated (coefficients of determination are given in Table XV) with GCP. Although high values of the coefficients of determination may partly be due to the small number of observations, the analysis revealed a high positive correlation between the percent vegetation cover represented in terms of GCP and the vegetation indices. The LEI had the highest correlation (correlation coefficient = 0.98) with GCP. Analysis of the slopes of the vegetation index curves (change in vegetation index for unit change in GCP) showed that when GCP was below 20, the NDVI curve had highest slopes. For higher GCP values the HU was most sensitive vegetation index. Since EH was seen to remain sensitive to changes in GCP over a wider range of GCP values, it would be the most suitable vegetation index for monitoring the restocking over clearcut forest lands. Table XTJI. Confusion matrix based on cluster groups of the reforestation areas Assumed Correct Classification 1 2 3 4 Total c 1 357 17 0 0 374 1 G u r 2 106 586 0 0 692 s 0 t u 3 0 91 340 0 431 e p r s 4 0 0 3 521 524 T o t a l 463 694 343 521 2021 Overall or diagonal accuracy = 88.82 percent Table XIV. Radiometric calibration coefficients for TM data TM Band Number Gaina Bias" 1 0.6024313569 -0.1519999981 2 0.1175098062 -0.2839999914 3 • 0.8057647347 -1.1699999570 4 0.8145489693 -1.5099999900 5 1.1080784276 -0.3700000048 7 0.0569803901 -0.1499999911 SOURCE: Radiometric ancillary record received from CCRS with the TM image digital data. "In w m 2 ster1 um'1 per DN b In w m'2 ster1 unr1 Table XV. Coefficients of determination between vegetation indices and GCP Vegetation Coefficient of Index Determination' LEI 0.9705 RVI 0.9440 NDVI 0.9423 "There is no exact statistic to compare correlation coefficient for testing the significance of non linear correlation. However, as a practical procedure, tabulated r (positive square root of coefficient of determination) values for (m - 1) variables and (n - m + 1) degrees of freedom, where m is the number of coefficients being estimated and n is the total number of observations, can be used as a measure of comparison (Draper and Smith, 1981). In the present case, tabulated r for 2 variables and 13 degrees of freedom is equal to 0.608, which indicates that each of the three vegetation indices has significant correlation with GCP. Figure 14. Infrared Index curve 1.0 IRI = 0.918 - 11.386/(GCP+35.992) X 0.8 — I R I X / / x 0.6: > 4 0.4 ( 1 1 1 1 1 ) 20 40 60 80 100 GROUND COVER PERCENT 60 Figure 15. Normalized Difference Vegetation Index curve Figure 16. Ratio Vegetation Index curve 62 CHAPTER 5 CONCLUSIONS Digital analysis of the remotely sensed data acquired through resource satellites is a powerful tool for studying the condition of forests. In this thesis, digital data obtained through the Landsat thematic mapper sensor system were investigated with regard to their information content for monitoring reforesting clearcuts. Techniques based on statistical separability measures were implemented to identify an optimum three band subset of the six band TM data for classifying reforesting areas into various restocking stages. Three commonly used vegetation indices were studied for quantitative assessment of restocking. Study of reforestation areas, situated near the Alex Fraser research Forest of the University of British Columbia, suggests the following: • Correlation analysis of the reflective thematic mapper bands indicates that the visible, near infrared, and shortwave infrared regions form three fundamentally independent spectral regions for remote sensing of vegetation. Principal component analysis also reveals that most of the information in six reflective thematic mapper bands is contained in the first three principal components. The first principal component is a measure of brightness, the second is a greenness type feature, and the third is a contrast between the short wave infrared and the other four bands. 63 • For discrimination of various restocking stages, the optimum three band subset comprises of thematic mapper bands 3 (the red band), 4 (the near infrared band), and 5 (the shortwave infrared I band). This combination makes use of spectral responses of vegetation in all three distinct spectral regions of the reflective thematic mapper data: the visible region characterized by high absorption due to leaf pigments; the near infrared region with high reflectance from the leaf cellular structure, and cellular constituents; and the short wave infrared region where reflectance is largely dominated by absorption due to leaf moisture content. ° Classification accuracy evaluation demonstrated that the reforestation areas can be classified into several restocking classes with approximately 90 percent accuracy. Average B-distances for all possible pairs of restocking classes under study indicated that the six band TM data would result in only marginal improvement in anticipated classification accuracy, confirming that a three band subset serves almost as well as the complete six band data in discriminating various restocking stages. • Correlation analysis indicates that forest stocking, measured in terms of ground cover percent, has high correlation with the vegetation indices. Infrared index was determined to be the most suitable vegetation index for quantitative monitoring of regrowth. Results of the study clearly support the research hypotheses. However, these findings should be regarded preliminary due to the small number of clearcuts examined. The results of this study were specific to ecosystem type, 64 species, and absence of slope and it would be premature to make extensive generalizations. Further investigations are necessary to evaluate thematic mapper data for temporal monitoring of reforestation. Additional research is also necessary to investigate the effects of topography on classification accuracies, and the performance of vegetation indices for quantitative assessment of vegetation. The classification of reforestation according to species composition also needs to be studied. 65 LITERATURE CITED Ahem, F. J., and P. D. Archibald, 1986. Thematic Mapper Information Contents About Canadian Forests: Early Results from Across the Country. Proc. 10th Can. Symp. on Remote Sens. :683-697. Ajai, D. S. Kamat, G. S. Chaturvedi, A. K. Singh, and S. K. Sinha, 1983. Spectral Assessment of Leaf Area Index, Chlorophyll Content, and Biomass of Chickpea. PE&RS, 49:1721-1727. Aldrich, R. C , 1975. Detecting Forest Disturbances in a Forest Environment. PE&RS, 41(l):39-48. Anon., 1987a. A Field Guide for Identification and Interpretation of Ecosystems, Province of British Columbia, Ministry of Forests and Lands, Williams Lake, B.C. Anon., 1987b. MOSAICS Technical Information Package. Canada Centre for Remote Sensing, Ministry of Energy, Mines and Resources, Ottawa, Canada, 18p. Anuta, P. E., L. A. Bartolucci, M. E. Dean, D. F. Lozano, E. Malaret, C. D. McGillem, J. A. Valdes, and C. R. Valenzuela, 1984. Landsat -4 MSS and TM Data Quality and Information Content Analysis. LEEE Trans, on Geosci. and Remote Sens., GE-22(3):222-236. Archibald, P. D., and F. J. Ahem, 1985. A Statistical Analysis of Forest Harvest Depletion Mapping Accuracy Using Landsat MSS Data. Intl. Symp. Remote Sens. Environ., Proc, 19th:697-709. Bansal, A. K , 1986. Landuse Mapping and Monitoring Using Landsat Data. Indian Forester, 112(1):6-12. Bansal, A. K , and P. A. Murtha, 1988. Use of Normal-Color Aerial Photographs in Collection of Data for Reforestation Assessment. Indian Forester (in press). Bhattacharyya, A., 1942. Proc. Ind. Sc. Congress, Quoted in A. Bhattacharyya, 1943. On a Measure of Divergence Between Two Statistical Populations Defined by Their Probability Distributions. Bull. Calcutta Math. Soc, 35:99-109. Bryant, E., A. G. Dodge, and M. J. E. Eger, 1979. Small Forest Cutting Mapped with Landsat Digital Data. Intl. Sypm. on Remote Sens, of Environ., Proc, 13th:971-981. Butera, M. K., 1986. A Correlation and Regression Analysis of Percent Canopy Closure Versus TMS Spectral Response for Selected Forest Sites in the San Juan National Forest, Colorado. IEEE Trans, on Geosci. and Remote Sens., GE-24(1):122-129. 66 Card, D. H., and G. L. Angelici, 1983. A Minicomputer Based Software System for the Selection of Optimal Subsets of Thematic Mapper Channels. Int. Geosci. and Remote Sens. Symp., Proc, Vol. 2, TP-4:6.1-6 Casella, G., 1983. Leverage and regression through the origin. American Statistician, 37(2):147-155. Cattell, R. B., 1966. The Scree Test for the Number of Factors. Multivariate Behavioral Research, l(2):245-276. Chance, J. E., 1981. Crop Identification and Leaf Area Index Calculations with Landsat Multi-Temporal Data. Intl. J. of Remote Sens., 2:1-14. Chavez, P. S. (Jr.), 1984. Digital Processing Techniques for Image Mapping with Landsat TM and SPOT Simulator Data. Intl. Symp. on Remote Sens, of Environ., Proc, 18th: 101-116. Chavez, P. S. (Jr.), G. L. Berlin, and L. B. Sowers, 1982. Statistical Method for Selecting Landsat MSS Ratios. J. of Applied Photo. Eng., 8(l):23-30. Choudhury, B. J., 1987. Relationship Between Vegetation Indices, Radiation Absorption and Net Photosynthesis Evaluated by a Sensitivity Analysis. Remote Sens, of Environ., 22:209-233. Colwell, J. E., 1974. Vegetation Canopy Reflectance. Remote Sens, of Environ., 3:175-183. Crapper, P. F., 1980. Errors Incurred in Estimating an Area of Uniform Land Cover Using Landsat. PE&RS, 46(10):1295-1301. Crist, E. P., and R. C. Cione, 1984. A Physically Based Transformation of TM Data - the Tasseled Cap. IEEE Trans, on Geosci. and Remote Sens., GE-22(3):256-263. Curran, P. J., 1980. Multispectral Remote Sensing of Vegetation Amount. Progress in Physical Geography, 4(3):315-341. Curran, P. J., 1983. Multispectral Remote Sensing for the Estimation of Green Leaf Area Index. Phil. Trans. R. Soc. London, A 309:257-270. Curran, P. J., and E. J. Milton, 1983. The Relationship between the Chlorophyll Concentration, LAI and Reflectance of a Simple Vegetation Canopy. Intl. J. of Remote Sens., 492):247-255. Curran, P. J, and H. D. Williamson, 1987. GLAI Estimation Using Measurements of Red, Near Infrared, and Middle Infrared Radiance. PE&RS, 53(2):181-186. DeGloria, S. D., 1984. Spectral Variability of Landsat-4 TM and MSS Data for Selected Crop and Forest Cover Types. LEEE Trans, on Geosci. and Remote Sens., GE-22(3):303-311. 67 Derenyi., E., D. Carlin, M. Dillon, D. Morgan, L. Peters, and R. Yardani, 1984. Change Detection Projects in New Brunswick. Proc. 9th Can. Symp. on Remote Sens.:799-806. Dillon, W. R., and M. Goldstein, 1984. Multivariate Analysis: Methods and Applications. John Wiley & Sons, Inc., New York, 587 p. Draper, N. R., and H. Smith, 1981. Applied Regression Analysis. 2nd ed., John Wiley & Sons, Inc., New York, 709 p. Dumoulin, F., 1985. Using Energy Measures for the Segmentation of Forest Scenes. Master's thesis, Univ.of British Columbia, Dept. of Forestry, Remote Sensing, 113 p. Elvidge, C. D., and R. J. P.Lyon, 1985. Estimation of the Vegetation Contribution to the 1.65/2.22 um Ratio in Airborne Thematic Mapper Imagery of the Virginia Range, Nevada. Intl. J. of Remote Sens., 6(1):75-88. Freden, S. C , and F. Gordon (Jr.), 1983. Landsat Satellites, pp. 517-570. In Manual of Remote Sensing, Vol. I, ed. R. N. Colwell, Am. Soc. of Photogrammetry, Falls Church, Virginia, USA, 1232 p. Friedel, J. P., and T. A. Fisher, 1987. MOSAICS - A System to Produce State-Of-The-Art Satellite Imagery for Resource Managers. Geocarto International, 3:5-12. Fu, K S., 1971. On Application of Pattern Recognition Techniques to Remote Sensing Problems. LARS Technical Note, TR-EE 71-13, Purdue Univ., West Lafayette, Indiana, USA, 91 p. Fu, K S., and C. H. Chen, 1965. Sequential Decisions: Pattern Recognition and Machine Learning. Tech. Report, TR-EE 65-5, School of Elec. Eng., Purdue Univ., West Lafayette, Indiana, USA. Fu, K S., D. A. Landgrebe, and T. L. Phillips, 1969. Information Processing of Remotely Sensed Agricultural Data. Proc. of the LEEE, 57(4):639-653. Fu, K S., P. J. Min, and T. J. Lee, 1970. Feature Selection in Pattern Recognition. LEEE Trans, on Systems Science and Cybernetics, SSC-6(l):33-39. Fung, T., and E. LeDrew, 1987. Application of Principal Component Analysis to Change Detection. PE&RS, 53(12):1649-1658. Gates, D. M., 1970. Physical and Physiological Properties of Plants, pp. 224-252. In Remote Sensing with Special Reference to Agriculture and Forestry. National Academy of Sciences, Washington, D.C, 424 p. Gates, D. M., H. J. Keegan, J. C. Schleter, and V. R. Weidner, 1965. Spectral Properties of Plants. Applied Optics, 4(l):ll-20. 68 Gausman, H. W., 1974. Leaf Reflectance of Near Infrared. Photogrammetric Engineering, XL(2):183-191. Gausman, H. W., 1977. Reflectance of Leaf Components. Remote Sens, of Environ., 6:1-9. Gausman, H. W., D. E. Escobar, J. H. Everitt, A. J. Richardson, and R. R. Rodriguez, 1978. Distinguishing Succulent Plants from Crop and Woody Plants. PE&RS, 44(4):487-491. Goel, N. S., and R. L. Thompson, 1984. Inversion of Vegetation Canopy Reflectance Models for Estimating Agronomic Variables: Estimation of Leaf Area Index and Average Leaf Angle Using Measured Canopy Reflectances. Remote Sens, of Environ, 16:69-85. Graetz, R. D., and M. R. Gentle, 1982. The Relationships Between Reflectance in the Landsat Wavebands and the Composition of an Australian Semi-Arid Shrub Rangeland. PE&RS, 48(11):1721-1730. Grettenberg, T. L., 1963. Signal Selection in Communication and RADAR Systems. LEEE Trans, on Infor. Theory, IT-9:265-275. Gross, M. F., V. Klemas, and J. E. Levasseyr, 1986. Remote Sensing of Spartina anglica Biomass in Five French Salt Marshes. Intl. J. of Remote Sens., 7(5):657-664. Hoffer, R. M., 1978. Biological and Physical Considerations in Applying Computer analysis Techniques to Remote Sensor Data. pp. 227-289. In Remote Sensing: The Quantitative Approach, eds. P. H. Swain and S. M. Davis, McGraw-Hill International Book Company, New York, 273 p. Hopkins, P. F., A. L. Maclean, and T. M. Lillesand, 1988. Assessment of Thematic Mapper Imagery for Forestry Applications Under Lake States Conditions. PE&RS, 5491):61-68. Horler, D. N. H., and F. J. Ahern, 1986. Forestry Information Content of Thematic Mapper Data. Intl. J. of Remote Sens., 7:405-428. Ihse, M., and W. Graneli, 1985. Estimation of Reed (Phragmites australis) Biomass Through Spectral Reflectance Measurements. Biomass, 8:59-79. Irons, J. R., B. L. Markhan, R. F. Nelson, D. T. Toll, D. L. Williams, R. S. Latty, and M. L. Stauffer, 1985. The Effects of Spatial Resolution on the Classification Accuracy of TM Data. Intl. J. of Remote Sens., 6(8):1385-1403. Janza, F. J., 1975. Interaction Mechanisms, pp. 75-179. In Manual of Remote Sensing, Vol. I, ed. R. G. Reeves, Am. Soc. of Photogrammetry, Falls Church, Virginia, USA, 867 p. Jarvis, P. G., G. B. James, and J. J. Landsberg, 1976. Coniferous Forests, pp. 171-240. In Vegetation and the Atmosphere, Vol. LI, ed. J. L. Monteith, Academic Press, 439 p. 69 Johnson, G. R., and E. W. Barthmaier, 1979. Forest Stand Classification in Western Washington Using Landsat and Computer-Based Resource Data. Intl. Symp. on Remote Sens, of Environ., Proc, llth:1681-1696. Jordon, C. F., 1969. Derivation of Leaf Area Index from Quality of Light on the Forest Floor. Ecology, 50:663-666. Kailath, T., 1967. The Divergence and Bhattacharyya Distance Measures in Signal Selection. LEEE Trans, on Comm. Tech., COM-15(1):52-60. Kauth, R. J., and G. S. Thomas, 1976. The Tasselled Cap - a Graphic Description of the Spectral-Temporal Development of Agricultural Crops as Seen by Landsat. Machine Processing of Remote. Sens. Data, Symp. Proc, 6/29-7/2, LARS, Purdue Univ., West Lafayette, Indiana, USA. Knipling, E. B., 1970. Physical and Physiological Basis for the Reflectance of Visible and Near Infrared Radiation from Vegetation. Remote Sens, of Environ., 1:155-159. Kumar, R., 1977. Feature Selection and Sample Classification Algorithms of INPE. In Remote Sensing of Earth Resources, technical papers selected from the Sixth annual Remote Sensing of Earth resources Conference, Tennessee, Vol. VI, ed. F. Shahrokhi, 497 p. Lee, Y. J., 1975. Reliability of Estimating Clearcut and Uncut Mature Timber Areas Using Landsat Imagery. Proc. 3rd Can. Symp. on Remote Sens.:265-272. Lillesand, T. M., and R. W. Kiefer, 1987. Remote Sensing and Image Analysis, 2nd ed., John Wiley & Sons, Inc., New York, 721 p. L6pez-Garcia, M. J., V. Caselles, and J. Milia, 1985. Following the Microclimatic Variations Produced by Forest Fires. Signatures Spectrales D'objects en Teledetection, Colloque., Proc, LES ARCS, France, 3rd:481-484. Min, P. J., D. A. Landgrebe, and K S. Fu, 1968. On Feature Selection in Multiclass Pattern Recognition. Ann. Princeton Conf. on Information Sciences and Systems, 2nd:453-457. Murtha, P. A., and E. K. Watson, 1975. Mapping of Forest Clearcutting: South Vancouver Island from Landsat Imagery. Proc. 3rd Can. Symp. on Remote Sens.:257-263. Myers, R. H., 1986. Classical and Modern Regression with Applications. Duxbury Press, Boston, Massachusetts, 359 p. NASA, 1982. Landsat Data Users Notes. Geological Survey, EROS Data Centre, Sioux falls, South Dakota, Issue 23, 13 p. Oswald, E. T., 1974. Evaluation of ERTS Imagery for Vegetation Interpretation in B.C. The Can. Surveyor, 28(2):147-154. 70 Pearse, P. H., A. J. Lang, and K L. Todd 1986a. The Backlog of Unstocked Forest Land in British Columbia and Impact of Reforestation Programs. The Forest. Chron., 62(6):514-521. Pearse, P. H., A. J. Lang, and K L. Todd 1986b. Economic Priorities for Reforesting Unstocked Forest Land in British Columbia. The Forest. Chron., 62(6):522-528. Pearson, E. S., and H. 0. Hartley, 1966. Biometrika Tables for Statisticians, Vol. 1. 264 p. Pearson, R. L., and L. D. Miller, 1972. Remote Mapping of Standing Crop Biomass for Estimation of Productivity of the Shortgrass Praire. Intl. Symp. on Remote Sens, of Environ., Proc, 8th:1457-1381. Pojar, J., K Klinka, and D. V. Meidinger, 1987. Biogeoclimatic Ecosystem Classification in British Columbia. For. Ecol. Manage., 22:119-154. Rencz, A. N., 1985. Multitemporal Analysis of Landsat Imagery for Monitoring Forest Cutovers in Nova Scotia. Can. J. of Remote Sens., 11(2):188-194. Richardson, A. J., and C. L. Wiegand, 1977. Distinguishing Vegetation from Soil Background Information. PE&RS, 43(12):1541-1552. Rouse, J. W. (Jr.), R. H. Hass, J. A. Schell, and D. W. Deering, 1973. Monitoring Vegetation Systems in the Great Plains With ERTS. NASA SP 351, ERTS-1 Symp., Proc, 3rd:309-317. Running, S. W., D. L. Peterson, M. A. Spanner, and K B. Teuber, 1986. Remote Sensing of Coniferous Forest Leaf Area. Ecology, 67(l):273-276. Shimabukuro, Y. E., P. H. Filho, and J. R. DosSantos, 1980. Automatic Classification of Reforestation in Pine and Eucalyptus Using Landsat. PE&RS, 46(2):209-216. Sievers, J. and K. Kriebel, 1980. Report of the ISP Working Group VQ-9 "Spectral Signatures of Objects". Proc. of the XTV Congress of the Intl. Soc. for Photogrammetry, Hamburg. Singh, A., and A. Harisson, 1985. Standardized Principal Components. Intl. J. of Remote Sens., 6(6):883-896. Singh, K D., 1980. Forest Cover Classification With Landsat Data. Proc. of the Training Course on Applications of New Remote Sensing Techniques to Forest Resources, Resources Survey, FAO, Rome. Snedecor, G. W., and W. G. Cochran, 1980. Statistical Methods, 7th ed. Iowa State Univ. Press. 507 p. Swain, P. H., 1972. Pattern Recognition: A Basis for Remote Sensing Data Analysis. LARS Info. Note 111572, LARS, Purdue Univ., West Lafayette, Indiana, USA, 41 p. 71 Swain, P. H., 1978. Fundamentals of Pattern Recognition in Remote Sensing, pp. 136-187. In Remote Sensing: The Quantitative Approach, eds. P. H. Swain and S. M. Davis, McGraw-Hill International Book Company, New York, 396 p. Swain, P. H., T. V. Robertson, and A. G. Wacker, 1971. Comparison of the Divergence and B-distance in the Feature Selection. LARS Information Note 020871, LARS, Purdue Univ., West Lafayette, Indiana, USA, 12 p. Swain, P. H., and R. C. King, 1973. Two Effective Feature Selection Criteria for Multi-Spectral Remote Sensing. LARS Info. Note 042673, LARS, Purdue Univ., West Lafayette, Indiana, USA, 5 p. Tassan, S., 1987. Evaluation of the Potential of the Thematic Mapper for Marine Application. Intl. J. of Remote Sens., 8(10):1455-1478. Townshend, J. R. G., J. R. Gayler, J. R. Hardy, M. J. Jackson, and J. R. Baker, 1983. Preliminary Analysis of Landsat-4 Thematic Mapper Products. Intl. J. of Remote Sens., 4(4):817-828. Tucker, C. J., 1978a. A Comparison of Satellite Sensor Bands for Vegetation Monitoring. PE&RS, 44(11):1369-1380. Tucker, C. J., 1978b. Are Two Photographic Infrared Sensors Required. PE&RS, 44(3):289-295. Tucker, C. J., 1979. Red and Photographic Lafrared Linear Combinations for Monitoring Vegetation. Remote Sens, of Environ., 8:127-150. Vickers, H., and R. J. Brown, 1986. Comparison of Landsat Thematic Mapper and Multispectral Scanner Information Content for Agricultural Applications in Western Canada. Proc. 10th Can. symp. on Remote Sens. :623-636. Wardley, N. W., and P. J. Curran, 1984. The Estimation of Green Leaf Area Index from Remotely Sensed Airborne Multispectral Scanner Data. Intl. J. Of Remote sens., 5(4):671-679. Weisberg, S., 1985. Applied Linear Regression. 2nd ed. John Wiley & Sons, Inc., New York, 324 p. Werle, D., Y. J. Lee, and R. J. Brown, 1986. The Use of Multi- Spectral and RADAR Remotely Sensed Data for Monitoring Forest Clearcuts and Regeneration Sites on Vancouver Island. Proc. 10th Can. Symp. on Remote Sens. :319-326. Wilkinson, L., 1988a. SYGRAPH. Evanston, IL, Systat Inc., 922 p. Wilkinson, L., 1988b. SYSTAT: The System for Statistics. Evanston, IL, Systat Inc., 822 p. 72 Williams, D. L., J. R. Irons, B. L. Markhan, R. F. Nelson, D. L. Toll, R. S. Latty, and M. L. Stauffer, 1984. A Statistical Evaluation of the Advantages of Landsat TM Data in Comparison to MSS Data. IEEE Trans, on Geosci. and Remote Sens., GE-22(3):294-301. Wilson, R. O., and P. T. Tueller, 1987. Aerial and Ground Spectral Characteristics of Rangeland Plant Communities in Nevada. Remote Sens, of Environ., 23:177-191. Wooley, J. T., 1971. Reflectance and Transmittance of Light by Leaves. Plant Physiology, 47:656-662. Zsilinszky, V., 1973. Take a Second Look. The Forest. Chron., 49(6):253-256. 


Citation Scheme:


Citations by CSL (citeproc-js)

Usage Statistics



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"
                            async >
IIIF logo Our image viewer uses the IIIF 2.0 standard. To load this item in other compatible viewers, use this url:


Related Items