@prefix vivo: . @prefix edm: . @prefix ns0: . @prefix dcterms: . @prefix skos: . vivo:departmentOrSchool "Arts, Faculty of"@en, "Geography, Department of"@en ; edm:dataProvider "DSpace"@en ; ns0:degreeCampus "UBCV"@en ; dcterms:creator "Abalharth, Mahdi Hadi"@en ; dcterms:issued "2013-09-26T15:52:18Z"@en, "2013"@en ; vivo:relatedDegree "Master of Science - MSc"@en ; ns0:degreeGrantor "University of British Columbia"@en ; dcterms:description "In-stream woods significantly influence watershed hydrology, flow regime, channel morphology and stability, and processes in streams. Consequently, in-stream woods play a major role in the existence and conservation of riparian and aquatic ecosystems. In this thesis, I attempt to detect and quantify LWD in stream channels using a remote sensing method, LiDAR, in conjunction with the traditional fieldwork. To the best of my knowledge, LiDAR-based analysis has not been used to study woods in stream channels. I, initially, attempted to re-apply advanced medical image processing and segmentation techniques on the LiDAR intensity images in order to confine the LiDAR terrain-based analysis to the stream channel networks, optimizing time and computing resources. The results exhibited significant image enhancement and accurate segmentation in certain regions; however, an automatic and a unified framework to delineate the stream channel networks, across different scales and spatial locations, is still required. LiDAR-based analysis demonstrated a more comprehensive solution for detecting in-stream woods in relation to the fieldwork through a high rate of commission and a low rate of omission. The filtered approach predicted the presence of 95% of fieldwork-reported in-stream woods, highlighting a 5% rate of omission, but with 25% rate of commission indicated by the identification of at least 15 new LWD locations that were not initially reported by the field crew. The non-filtered approach identified 87% of field-reported LWD, highlighting a 13% rate of omission and, similar to the filtered approach, a %25 rate of commission. Overall, the non-filtered and the filtered LiDAR showed fairly accurate predictions for in-stream woods’ dimensional measurements (length, width, and height) with respect to the field data. However, the filtered approach showed better dimension estimation of in-stream woods compared to the unfiltered LiDAR. Although a margin of error existed for fieldwork and LiDAR methods, a careful examination of orthophotos showed that LiDAR results were more accurate than the Laser Range Finder (LRF) used in the field."@en ; edm:aggregatedCHO "https://circle.library.ubc.ca/rest/handle/2429/45128?expand=metadata"@en ; skos:note """USING ?LIDAR ?TO ?DETECT ?IN-??STREAM ?WOODS: ?A ?SCALED ?APPROACH ? by Mahdi Hadi Abalharth B.S., California State University, Long Beach, 2005 A THESIS SUBMITTED IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF MASTER OF SCIENCE in The Faculty of Graduate and Postdoctoral Studies (Geography) THE UNIVERSITY OF BRITISH COLUMBIA (Vancouver) September 2013 ? Mahdi Hadi Abalharth, 2013 ? ? ii ?Abstract In-stream woods significantly influence watershed hydrology, flow regime, channel morphology and stability, and processes in streams. Consequently, in-stream woods play a major role in the existence and conservation of riparian and aquatic ecosystems. In this thesis, I attempt to detect and quantify LWD in stream channels using a remote sensing method, LiDAR, in conjunction with the traditional fieldwork. To the best of my knowledge, LiDAR-based analysis has not been used to study woods in stream channels. I, initially, attempted to re-apply advanced medical image processing and segmentation techniques on the LiDAR intensity images in order to confine the LiDAR terrain-based analysis to the stream channel networks, optimizing time and computing resources. The results exhibited significant image enhancement and accurate segmentation in certain regions; however, an automatic and a unified framework to delineate the stream channel networks, across different scales and spatial locations, is still required. LiDAR-based analysis demonstrated a more comprehensive solution for detecting in-stream woods in relation to the fieldwork through a high rate of commission and a low rate of omission. The filtered approach predicted the presence of 95% of fieldwork-reported in-stream woods, highlighting a 5% rate of omission, but with 25% rate of commission indicated by the identification of at least 15 new LWD locations that were not initially reported by the field crew. The non-filtered approach identified 87% of field-reported LWD, highlighting a 13% rate of omission and, similar to the filtered approach, a %25 rate of commission. Overall, the non-filtered and the filtered LiDAR showed fairly accurate predictions for in-stream woods? dimensional measurements (length, width, and height) with respect to the field data. However, the filtered approach showed better dimension estimation of in-stream woods compared to the unfiltered LiDAR. Although a margin of error existed for fieldwork and LiDAR methods, a careful examination of orthophotos showed that LiDAR results were more accurate than the Laser Range Finder (LRF) used in the field. ? ? iii ?Preface I am the sole owner of all the Figures in this thesis except Figures 1 and 6 that are used with permission from co-owner, Abdul Aziz Alkanhal. ? ? iv ?Table of Contents Abstract ?....................................................................................................................................... ?ii ?Preface ?....................................................................................................................................... ?iii ?Table ?of ?Contents ?.................................................................................................................... ?iv ?List ?of ?Tables ?............................................................................................................................ ?vi ?List ?of ?Figures ?.......................................................................................................................... ?vii ?Acknowledgements ?................................................................................................................ ?ix ?Dedication ?.................................................................................................................................. ?x ?1. ? Introduction ?...................................................................................................................... ?1 ?1.1 ? Wood ?In ?Streams ?..................................................................................................................... ?1 ?1.2 ? LiDAR ?Technology ?.................................................................................................................. ?4 ?1.3 ? Remote ?Sensing ?And ?Image ?Enhancement ?Literature ?............................................... ?7 ?1.4 ? Study ?Area ?................................................................................................................................. ?9 ?1.5 ? Data ?.......................................................................................................................................... ?12 ?2. ? Methods ?............................................................................................................................ ?14 ?2.1 ? Image ?Analysis ?...................................................................................................................... ?14 ?2.1.1 ? Data ?Preparation ?........................................................................................................................... ?14 ?2.1.2 ? Non-??Linear ?Diffusion ?................................................................................................................... ?15 ?2.1.3 ? Stream ?Channel ?Segmentation ?Techniques ?....................................................................... ?17 ?2.1.3.1 ? Vessel ?Enhancement ?Filter ?................................................................................................................ ?17 ?2.1.3.2 ? Region-??Growing ?..................................................................................................................................... ?19 ?2.1.4 ? Logjam ?Detection ?Using ?Spectral ?Signature ?...................................................................... ?19 ?2.2 ? Terrain ?Analysis ?................................................................................................................... ?20 ?2.2.1 ? Producing ?Digital ?Terrain ?Model ?(DTM) ?............................................................................. ?21 ?2.2.2 ? Producing ?Digital ?Surface ?Model ?(DSM) ?.............................................................................. ?24 ?2.2.3 ? Logjam ?Volume ?Detection ?......................................................................................................... ?26 ?3. ? Results ?& ?Discussion ?.................................................................................................... ?29 ?3.1 ? Image ?Analysis ?Results ?...................................................................................................... ?29 ?3.1.1 ? Image ?Restoration ?And ?Enhancement ?................................................................................. ?29 ?3.1.2 ? Non-??Linear ?Diffusion: ?.................................................................................................................. ?30 ?3.1.3 ? Enhancement ?Filter ?...................................................................................................................... ?32 ?3.1.4 ? Region-??Growing ?............................................................................................................................. ?33 ?3.1.5 ? Logjam ?Identification ?.................................................................................................................. ?34 ?3.2 ? DTM ?Comparisons ?............................................................................................................... ?35 ?3.3 ? DSM ?Results ?........................................................................................................................... ?37 ?3.3.1 ? Factors ?In ?LWD ?Detection ?......................................................................................................... ?37 ?3.3.2 ? Unfiltered ?LiDAR ?Returns ?.......................................................................................................... ?38 ?3.3.3 ? Filtered ?LiDAR ?Returns ?.............................................................................................................. ?42 ?4. ? Limitations ?...................................................................................................................... ?44 ?5. ? Future ?Research ?............................................................................................................ ?45 ?6. ? Conclusion ?....................................................................................................................... ?46 ? ? ? v ?References ?.............................................................................................................................. ?47 ? ? ? ? ? vi ?List of Tables ?Table 1: Basic LiDAR terminology. ................................................................................... 6 Table 2: Eigenvalues for different options of tubular structures. A dark tubular structure indicates a low change of color along the shape and a high positive value for the change of color across that same shape. ................................................................... 18 Table 3: The Standard LiDAR Point Classification of the American Society for Photogrammetry and Remote Sensing (ASPRS)(Graham, 2012). ........................... 21 Table 4: The definitions and values considered for the user-parameters of the MCC algorithm. .................................................................................................................. 23 Table 5: Image features and their corresponding experimental intensity values (0-255). 34 Table 6: Number of in-stream woods identified by each of the different methods attempted. The omission and commission errors are based on the fieldwork counts. The overall improvement is the result of subtracting the omission error from the commission error. ..................................................................................................... 35 Table 7: Height comparison of selected LWD midpoints between Multicurvature Classification algorithm (MCC) and the point classification. .................................. 36 ? ? vii ?List of Figures Figure 1: Illustration of how multiple-return LiDAR works. Multiple returns, from canopy, tree branches, shrubs, and ground, are collected for every laser beam. The entire point returns represent the point cloud. ............................................................ 7 Figure 2: The study area included the Wildhay River and Moberly Creek in Hinton, Alberta. ...................................................................................................................... 10 Figure 3: The study area of the lower Elwha River in Washington State, USA. .............. 10 Figure 4: Eigen values principle directions. ?1 represents the gradient (color) change along the tubular structure while ?2 represents the gradient change across the tubular structure. .................................................................................................................... 18 Figure 5: From the entire point cloud of the Wildhay River, yellow points are the extracted point-returns that were classified as ground returns. In general, ground returns were located on water surface, floodplain, and in between gaps of canopy and LWD. .................................................................................................................. 24 Figure 6: (A) shows how high vegetation could hinder the detection of the LWD boundaries. (B) illustrates how feasible is the delineation after excluding point-returns higher than 3 m above ground level. ............................................................. 25 Figure 7: (A) is a hillshade illustration of an ideal scenario of mid-channel LWD. (B) is the intensity image of the same spatial location showing bright pixels for LWD. ... 27 Figure 8: (A) is a Height Model showing the result of a raster math calculation, subtracting the DTM from DSM model. (B) is a reference hillshade image based on the DSM. ................................................................................................................... 27 Figure 9: (A) is the initial image exhibiting noise and missing data. (B) is the restored image. ........................................................................................................................ 30 Figure 10: (A) shows the impact of non-linear diffusion that intuitively described in-stream woods as part of the channel. (B) illustrates the lack of traditional Gaussian smoothing to include such LWD as part of the channel. .......................................... 31 Figure 11: The enhancement filter applied to the Wildhay River merging with Moberly Creek. (A) is the original image before the enhancement filter. (B) is the image after the enhancement filter. .............................................................................................. 33 Figure 12: An illustration of a segment of the Wildhay River channel enhancement filter. (A) is the original image. (B) is the output of enhancement filter showing scattered artifacts around the stream channel. (C) is the channel segmentation overlaid on the original image. .......................................................................................................... 33 Figure 13: An illustration of a channel segmentation using region-growing technique. (A) is the original image. (B) is the region-growing boundaries results overlaid the original image. .......................................................................................................... 34 Figure 14: The spatially spread LWD-midpoints on which candidate DTMs were compared. .................................................................................................................. 37 ? ? viii ?Figure 15: Comparisons of in-stream woods dimensions across the three types of vegetation obstructions. (A) column compares the field data to LiDAR measurements before the filtering while (B) column compares measurements after the filtering. ............................................................................................................... 41 Figure 16: The impact of filtering LiDAR point cloud on LWD visibility and boundary delineation. The circles on (A) illustrated 2 locations of in-stream woods that were not visible in (B) before filtering the LiDAR points. ................................................ 42 ? ? ? ix ?Acknowledgements I offer my sincere gratitude to my supervisors, Marwan Hassan and Brian Klinkenberg, who made it a pleasant experience and my fellow students and lab mates at the UBC Geography department who provided a great working environment and continuous support throughout this journey. I am grateful for Nicolas Coops for the Remote Sensing experience, LiDAR in particular, and for his support in SilviLaser 2012 conference participation. I thank Rafeef Abugharbieh for providing a remarkable experience in Fundamentals of Visual Computing and engaging that knowledge into a hydrology context. I would like to extend my appreciation to Rich McCleary, a former PhD student of Marwan Hassan, for general help and guidance through the work that has been accomplished on Hinton Dataset and arranging for my fieldwork. I would like also to thank Hinton Wood Products Ltd. and Glenn Buckmaster, in particular, for providing the initial dataset for this project. Thanks also extended to the Puget Sound LIDAR Consortium and University of Washington, particularly Vivian Leung who is a graduate student at the Earth and Space Sciences Department, for offering a higher resolution LiDAR and field data, respectively. I am also grateful for Guy Gilboa for his Matlab implementation of the Perona and Malik algorithm for non-linear diffusion. Special thanks to my employer, Saudi Aramco, as this would not be a reality without their financial and mental support through out the years. ? ? x ?Dedication To my parents and wonderful family ? 1 ?1. Introduction 1.1 Wood In Streams Forests, through the abundant supply of wood into the stream channel network, significantly influence watershed hydrology, flow regime, channel morphology, and stream ecology (e.g., Hassan et al., 2005). Such effects are particularly the result of the in-stream accumulation of large woody debris (LWD) and channel-margin trees. LWD accumulations reduce sediment transport efficiency through decreasing the probability of bed particle entrainment and reducing the average travel distance of entrained particles (Beschta, 1979; Bilby, 1981; Heede, 1972, 1975; Heede et al., 1972; Marston, 1982; Megahan, 1982; Nakamura and Swanson, 1993; Swanson and Lienkaemper, 1978; Thompson, 1995). Despite their size, in-stream woods act as roughness elements (Gurnell et al., 2002) and their quantity, position, and orientation have different impact on flow resistance, patterns, and water surface profile (Gippel et al., 1996; Young, 1991). Faustini and Jones (2003) concluded that in-stream woods create structures, such as steps and pools, which may trap sediment and reduce their transport, causing high variation in bed elevation, water depth, and particle size. Not only do in-stream woods constitute a source of food for aquatic and riparian plants and animals, but also they play a major role in sustaining water quality and providing refuge habitat that protects organisms during pollution and high flow incidents (Gurnell et al., 2002). They provide vital contributions to the health of streams, rivers, and oceans (Maser and Sedell, 1994). Intensive research on in-stream woods started around the late 1970?s when fish biologists noticed it was a major factor in habitat quality?prior to that, in-stream woods were intentionally removed from channel networks assuming that would improve fish passage (Hyatt and Naiman, 2001). As woods in-stream stabilize channel bed and banks (Smith et al., 1993; Scrivener, 1987), they create diverse habitat (Crispin et al., 1993), and protect fish from high flows and predators (Martin et al., 1986). The abundance of such woods was correlated with higher densities of juvenile salmonids in winter (Murphy et al., 1984) and with more diverse types of pool and riffle (Lonzarich, ? 2 ?1994). Through regulating sediment and organic matters transport, in-channel woods provide temporal and spatial source of nutrients for aquatic habitat (Gurnell et al., 1995). In an attempt to minimize the effect of urbanization on stream channels near Vancouver, British Columbia, the introduction of in-stream woods was identified as the most significant strategy for stream restoration and was the most important factor for improving the physical elements of fish habitat (Finkenbine et al., 2000). According to Hyatt and Naiman (2001), woods in streams also provide critical functions in river ecosystems, such as sediment and nutrient retention, improving fish habitat, and stabilizing colonization sites for incipient floodplain vegetation. The complex structure and pattern of in-stream woods create diverse habitat that can shelter and sustain a wide range of organisms at different stages of their lives (Gurnell et al., 2002). As a proactive measure, the Oregon Department of Forestry encourage in-stream woods delivery to fish bearing streams by mandating forest buffers in certain headwater streams (Benda et al., 2005). In headwater streams, the size and physical properties of LWD exerts a great geomorphic influence on the channel?s longitudinal profile (Faustini and Jones, 2003). Typically, river size controls most of in-stream wood physical characteristics such as wood supply, mobility, entrainment, deposition, accumulation, and temporal variability (e.g., Gurnell et al., 2002; Hassan et al., 2005). Because of their smaller widths, headwater streams do not often transport large woods (e.g., Millard and Region, 2001), causing large buildups of in-stream woods (Hogan et al., 1998; Jackson and Sturm, 2002; May and Gresswell, 2003) and sediments (Benda et al., 2005). Such high sediment and wood storage in headwater streams lead to temporal variation in channel morphology due to debris flows and flash floods, and thereby adding complexities to the classification of those channels (Benda et al., 2005). The impact and functionality of in-stream woods changes with the size of the stream channel. For instance, in small channels, LWD may span the whole channel creating bridges that have little or no impact on the channel. In medium-size channels, woods can block part of or the entire channel exerting the greatest influence on landscape, ecology, ? 3 ?and geomorphology. However, the functionality of wood declines in large channels, as logjams are too small to block the entire channel. Overall, the hydraulic impact on the landscape varies proportionally with the wood blockage ratio1 across the stream channel. In a study examining the impact of in-stream woods on patterns of sediment deposition on forested floodplain in England, Jeffries et al. (2003) concluded that woods significantly influence channel hydrology and hydraulics, and thereby leading to increased overbank discharge of water and sediment. They also observed that floodplains impacted by in-stream woods had greater and more varied amounts of sediment depositions. Understanding the spatial pattern of in-stream wood, geomorphic processes and channel characteristics is fundamental for studies of landscape evolution, aquatic ecology, conservation biology, and river restoration. Channel reaches define a useful scale over which to relate stream morphology to channel processes, wood dynamics, response potential, and habitat characteristics. Forest management is increasingly tied to the maintenance of fish habitat, in part because different types of channels produce different types and quality of habitat for different species. Consequently, the ability to predict the spatial and temporal patterns of in-stream woods and channel morphology would be valuable for a wide range of applications. Traditionally, digital remote sensing and fieldwork have been used to study woods in streams. Traditional digital remote sensing primarily utilizes the spectral signatures to identify objects of interest and, hence, is less helpful in canopy-obstructed channels and also is not able to predict logjam heights. Fieldwork can provide better size predictions despite canopy obstruction, but is considered expensive, unrepeatable, time consuming, and challenging when certain parts of the channel network are inaccessible. LiDAR, as a new remote sensing technology, has the advantage of providing accurate 3-D analysis allowing for better feature detection and sizing. ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ?1 ?The ?percentage ?of ?the ?channel ?blocked ?by ?woods. ?The ?higher ?the ?percentage, ?the ?higher ?the ?chances ?of ?water ?overflow. ? ? 4 ?Narrowing the entire LiDAR coverage area to an appropriate scale is very useful technique in studies concerning stream channel networks. Typically, LiDAR datasets cover large areas of various topographies in which stream channel networks constitute a small portion. Hence, restricting data analysis to only the extent of the channel networks is critically important in optimizing time and computing resources. A common approach to delineate the channel networks is the use of shapefiles, a geospatial vector data format that can describe the boundaries of channel networks through digitization of orthophotos. Temporally, shapefiles can become obsolete and must be recreated to depict an up-to-date channel banks? a cumbersome and time-consuming task. Therefore, there is a great demand for an automated tool that accurately delineate the stream channel networks in a timely manner, in conjunction with the traditional shapefiles. The objective of this thesis is to identify methods that could be used to automate the boundary-delineation of stream channel networks, and consequently, explore the use of the LiDAR in identifying locations of in-stream woods and estimating the volume of such woods in the channel network. The study also attempts to examine the impact of channel size and canopy obstruction on the detection process. To achieve such goals, I obtained and analyzed LiDAR and field data from two study reaches; Wildhay River in Hinton, Alberta, and the lower Elwha River in Washington State. In the remaining part of my thesis, I use the terms in-stream woods, LWD, and logjams interchangeably. 1.2 LiDAR Technology LiDAR (Light Detection And Ranging) is an active remote sensing technology that emits its own laser radiations and measures both the time and backscattered light of the reflecting objects. Depending on the application, the laser radiations can be in the ultraviolet, visible, or near infrared range of the electromagnetic spectrum. It is a fairly new and expensive technology that is equipped with sophisticated position and navigational systems capable of producing accurate measurements despite the frequent and random airplane movements, such as pitching, yawing, and rolling. For instance, the Inertial Measurement Unit (IMU) provides precise measurement of aircraft velocity, orientation, and gravitational forces. A Global Positioning System (GPS) provides ? 5 ?information about the time and XYZ location of the aircraft. Such precise information about the aircraft and the sensor are utilized to construct an accurate 3-D point cloud, a spatially organized set of LiDAR returns. Because of its accurate surface and terrain measurements, LiDAR technology has a wide-range of applications in archaeology, forestry, military, geology, hydrology, and soil science. There are two types of the LiDAR technology: (1) Airborne, on which the LiDAR instruments are installed on an aircraft and can be subdivided into two further categories: (a) Topographic, used to derive surface models for various earth topographies. (b) Bathymetric, capable of penetrating water and hence, can derive underwater surface models. (2) Terrestrial, which can be subdivided into: (a) Mobile, on which LiDAR instruments are installed on a moving platform, such as a car or train, and are typically used to analyze road infrastructure. (b) Static, where the instruments are in static location and are typically used in engineering and surveying applications. The topographic airborne LiDAR emits radiation in the infrared region of the electromagnetic spectrum whereas the bathymetric LiDAR works in the green region of the spectrum. Current airborne LiDAR is capable of emitting a pulse rate of 150-300 kilohertz (i.e., 150,000-300,000 pulses per second) and recent instruments can send pulses before receiving the previous pulses, allowing for a higher point density on the target. The accuracy of the LiDAR data is reported in the form of horizontal and vertical accuracies that are determined by factors such as the flight altitude, footprint size, beam divergence, pulse rate, point spacing, etc. The horizontal accuracy refers to the uncertainty in the X-Y coordinates of the LiDAR returns whereas the vertical accuracy refers to the uncertainty in the Z coordinate of the LiDAR returns. In Table 1, I describe common LiDAR terminology. Earlier versions of LiDAR data were delivered in XYZ file format, primarily storing the X-Y-Z coordinates of each LiDAR return; however, LAS is a newer file format that can store additional information such as intensity, point classification, return number, number of returns, etc. ? ? 6 ?Terminology Definition Beam Divergence An angular measure at which the pulse is emitted from LiDAR instrument and, consequently, the footprint size is determined. It is important as it also determines the pulse penetration in forests and highly vegetated areas. DEM A surface created from an elevation model (3-D point cloud) to represent any topography. It is often used as a generic name for DTM and DSM. DSM A surface created from the 3-D point cloud to represent features on or above ground, typically the top surface of features above ground. DTM A surface created from the 3-D point cloud to represent the bare-ground without objects like trees and buildings. Footprint The cross-sectional diameter at the target surface when hit by a pulse. The smaller the beam divergence, the smaller the footprint, which creates high energy at the footprint capable of penetrating forests and dense vegetation. Intensity Data The strength and energy of the pulse returned to the sensor. Normally, the energy decreases as the number of return increases per pulse. Number of Returns The total number of returns captured per pulse. Point Classification Information about the object that reflected the pulse. The data provider delivers it after post-processing of LiDAR data. Point Cloud The spatially organized LiDAR returns. Point Density The number of LiDAR returns per unit square. Point Spacing 1-dimensional parameter describing the point-to-point distance of LiDAR returns. It determines the (optimal) resolution of derived grids such as the raster surface, terrain, and Triangulated Irregular Network (TIN). Pulse Rate The number of emitted pulses per second by the LiDAR instruments. Return Number An emitted laser pulse can have up to 5 returns, nowadays. The return number is the order of returns per pulse. Hence, the first return will have a value of one for the return number and the second return will have a value of two, etc. RMS/RMSE Root Mean Square Error is a common accuracy measure for the horizontal and vertical accuracy. Table 1: Basic LiDAR terminology. Figure 1 is an illustration of an aircraft equipped with LiDAR instruments emitting laser pulses towards the ground. The first pulse return to the LiDAR sensor reflected off the top part of the tree canopy and the second return reflected from a lower tree branch. The third return bounced off shrubs below the tree and the last reflection returned from the ground (Figure 1). Spatially, each return or reflection represents a single point in the point cloud delivered by a LiDAR data provider. ? 7 ? Figure 1: Illustration of how multiple-return LiDAR works. Multiple returns, from canopy, tree branches, shrubs, and ground, are collected for every laser beam. The entire point returns represent the point cloud. 1.3 Remote Sensing And Image Enhancement Literature In this section, I provide a brief review of the relevant history of the applied techniques. First, there is a review of the earlier attempts using non-LiDAR remote sensing applications to supplement the traditional fieldwork. Next is a survey of recent LiDAR-based DEM attempts to detect and estimate fallen wood in forests. Lastly, a number of advanced image processing techniques are reviewed that enhance the quality and coarse resolution of the LiDAR intensity images? techniques that attempt to enhance the gray-scale images extracted from the LiDAR point cloud and images that are partly characterized by noise and missing data. In a fluvial context, remote sensing has been and is still an active research field for stream channel delineation and LWD detection. Marcus et al. (2003) used high-spatial resolution hyperspectral remote sensing to measure stream channel depth. They also used a matched filter algorithm in a supervised classification approach to detect logjams within channels, an interactive approach that requires user-defined spectral information of common LWD. ? 8 ?Smikrud et al. (2008) looked at LWD outside stream channels using high-resolution aerial photography and showed that segmentation of LWD should be based on contextual information, as LWD do not display a unique spectral signature in the visible spectrum in comparison to gravel and sand. They also suggested the need for user-interaction in delineating the stream channels due to the presence of prominent shadows from nearby trees, which obstruct the view in the visible range. Utilization of the near infrared region of the electromagnetic spectrum can overcome the tree shadow effect?this is an advantage that the near infrared region offers over the visible range (Smikrud et al., 2008). After LiDAR was introduced as a remote sensing technology, studies were carried out to see if it could be used to identify and estimate the shape and volume of fallen wood in various environments. For instance, Blanchard et al. (2011) attempted to delineate single downed wood utilizing an Object-Based Image Analysis (OBIA) approach and very dense LiDAR data, 5 cm vertical accuracy and 5-10 cm horizontal accuracy. Blanchard et al. (2011) produced a Triangulated Irregular Network (TIN) surface from the LiDAR data and consequently applied image segmentation and classification techniques to delineate the downed log objects. In a statistical model, Pesonen et al. (2008) predicted coarse woody debris (CWD) in non-disturbed forested landscapes by using variables obtained from the LiDAR height intensity metrics to construct a CWD regression model that can estimate the volume of dead downed trees. There have been few investigations into the semi-automatic identification of stream channels using remotely sensed images. Interestingly, stream channel segmentation is a similar task to vascular vessel segmentation except that the first may allow for more complex bifurcation and branching. Vascular diseases that affect vessels are common in old patients and require accurate quantification and visualization during interventional procedures (Zhou et al., 2007). Such demand for accurate information has enriched the literature of vessel enhancement and segmentation techniques. Because there has been more advanced research conducted in vessels segmentation than in stream channels, I attempt to apply this segmentation literature in the stream channel delineation context. ? 9 ? Recent medical research has exhibited major advances, particularly in image acquisition and processing. However, according to Dehkordi et al. (2011), vascular structure enhancements that target noise and background suppression, such as that proposed by Frangi et al. (1998), is of limited effect as noise and background structures similar to vessels might not be suppressed. Also, low-contrast segments of vessel structure are likely to be suppressed when removing background structures. These drawbacks may also apply to the situation with respect to stream channel enhancement techniques. Various vessel segmentation techniques, based on pattern recognition, model-based tracking and propagation, neural network, fuzzy, and artificial intelligence, have been developed (Dehkordi et al., 2011). These approaches, in general, share a common disadvantage as they are fixed-scale and unable to detect wide-ranges of vessels, and their extension to multiple-scales would lead to high computational requirements. Nonetheless, many semi-automated and training-based models for vessels and tube-structures exist. For example, the interactive program Live-Vessel for segmenting vessels in the retina display (Poon et al., 2007) is based on the Frangi et al. (1998) vessel enhancement filter. However, the methods used in Live-Vessel cannot handle complex vessel bifurcations. 1.4 Study Area The research was conducted on the Wildhay River and Moberly Creek in Hinton, Alberta (Figure 2), and the Elwha River in Washington State (Figure 3). In general, those study sites represented watershed areas with dense forest cover. The vegetation cover that is adjacent to the channels in both sites has a great impact on the visibility of channel banks in aerial imagery as well as on LiDAR-based terrain analysis. For instance, tall trees on channel banks create shadows in the visible region of the electromagnetic spectrum, especially if those images were not captured at noon, making interpretation of orthophotos much more difficult. That is, depending on the size of the shadow created, visibility issues arise due to the lack of illumination on targeted areas. Hence, both the channel boundaries and any shaded LWD may not be visible. ? 10 ? ?Figure 2: The study area included the Wildhay River and Moberly Creek in Hinton, Alberta. ?Figure 3: The study area of the lower Elwha River in Washington State, USA. The study area of the Wildhay watershed includes the Lower and Upper Foothills, sub-regions of the Foothills Natural Region, with headwaters in the Rocky Mountain Natural Region. The Lower Foothill sub-region has a mean annual precipitation and temperature of 464 mm and 3C?, respectively, with mean winter temperature of -7.8C? and 12.8C? in the summer (Beckingham et al., 1996). Elevation ranges from approximately 500-1150 m ? 11 ?and the area is dominated by aspen (Populous ? tremuloides), balsam poplar (Populous ?balsamifera), lodgepole pine (Pinus ?contorla), and white spruce (Picea ?glauca) (Strong 1992). The topography is generally rolling with ridges underlain by sandstone and shale along rocky mountain edge (Beckingham et al., 1996). The Upper Foothill sub-region has a mean annual precipitation and temperature of 538 mm and 3C?, respectively with mean winter temperature of -6C? and 11.5C? in the summer. The elevation ranges from approximately 900-1500 m (Corns and Anna, 1986) and the area is dominated by closed-canopy coniferous forest and lodgepole pine species in particular, and is also characterized by lack of aspen. The topography is best described by strongly rolling ridges along the rocky mountain edge with marine shales and non-marine sandstone bedrocks (Beckingham et al., 1996). The second study area is the Elwha River, which flows north for 72 km, with elevations ranging from 1372 m at the headwaters to sea level at the mouth. The climate is generally dry and warm in the summer and cool and wet in the winter (Duda et al., 2008). The mean annual precipitation of the lower basin is 1000 mm (Phillips, 1972) and the high discharge occurs in the fall and winter because of increased precipitation, and also in the spring due to snowmelt in the upper parts of the basin (Duda et al., 2008). Unconsolidated glacial till and alluvial deposits mostly characterizes the lower part of the river (Tabor, 1987). A series of alternating canyons and floodplains characterizes the geomorphology of the river (Pess et al., 2008), where the channel meanders between canyons depending on vegetation, LWD, and sediment (Latterell et al., 2006). The lower elevation forests in the river are mostly Douglas Fir with a mixture of western hemlock and western red cedar (Duda et al., 2008). Parts of both rivers, the Wildhay and Elwha, flow under-canopy, leading to further challenges in areas where the canopy is very dense. Because current LiDAR technology cannot penetrate dense and closed vegetation, a loss of information on terrain and features underneath such vegetation can occur. The volume of in-stream woods varies widely in both rivers. For instance, wood accumulations in the Elwha River ranges from single pieces of dead wood to collections of debris around 100x40 m2 in area. This ? 12 ?variability can add challenges for the field crews when attempting to gather comprehensive and accurate information about the actual dimensional measures of LWD. 1.5 Data For the study conducted in Hinton, Alberta, Hinton Wood Products (HWP), a division of West Fraser Mills Limited, provided the LiDAR data that cover an area of approximately 400 km2 (Figure 2). The delivered LiDAR dataset is considered raw point cloud data that has not been filtered or modified after the initial post-processing by data providers. This dataset was acquired through topographic airborne LiDAR with a flight altitude around 1,200 m above ground level on June 25, 2006 and with horizontal accuracy of 45 cm RMS and vertical accuracy of 30 cm RMS. The point spacing ranges from 0.65 to 0.78 m, equivalent to 1.6 to 2.4 point density, meaning there are 1.6-2.4 LiDAR returns per m2, and hence an output raster of 1-meter resolution would be acceptable. For validation purposes, HWP also delivered single-band orthophoto images of the same study area (Figure 2). Although there was 1-2 months gap between the acquisition times of the LiDAR data and orthophotos, I assumed the data were appropriate for validation because both datasets were taken during the summer, which was a low-flow season with little likelihood of in-stream wood movements. This dataset was originally collected to provide a common digital elevation database to enable the Government of Alberta to approve changes to the forest companies? harvest in an expedient manner. Government departments and agencies in the course of their normal work activities can also employ this data when a more accurate digital elevation model (DEM) is required. However, no ground-truth information of wood spatial location and size was available, and hence no validation of the LiDAR-derived results could be made. The second LiDAR dataset was provided by the Puget Sound LIDAR Consortium (PSLC), while the ground-truth information was provided by Vivian Leung (Personnel Communication, March 2013). The LiDAR and validation datasets were taken at approximately the same time in 2012. The point spacing ranges from 0.19 to 0.34 and the point density ranges between 8.9 and 27, meaning that there are 8.9-27 LiDAR returns ? 13 ?per m2, and hence an output raster of 0.3-m resolution is acceptable in the lower part of the Elwha river in Washington State, USA. The team that acquired the field data used a raft to reach the possible locations of major in-stream wood, and used a Laser Range Finder (LRF) device to estimate the length, width, and height of each recorded LWD. The LRF device is positioned at one-end of the LWD before it emits laser beams to the endpoint at the other side. Then, it automatically calculates the distance between the two ends. The field data contained 60 in-stream logjams varying in size and location with reference to the stream channel and vegetation cover. According to Leung (Vivian Leung, Personnel Communication, March 2013), some LWD locations were skipped because they were un-reachable, illustrating another challenge for fieldwork. A quick examination of Hinton and Elwha LiDAR datasets revealed that Elwha?s had higher resolution (average of 0.3-m), allowing for more precise measurements across various scales of stream channels and in-stream wood volumes. Nonetheless, working on two LiDAR datasets with varying resolution allowed for determining the level of details each dataset can provide and the smallest logjam that can be detected. My goals in the remaining chapters are to develop a tool capable of segmenting the boundaries of the Wildhay River and Moberly Creek from the LiDAR intensity images and compare the results to the orthophotos provided by HWP. This tool would automate the segmentation process, optimizing time and computing resources, when analyzing LiDAR data. This can be achieved through focusing the data analysis on the extent of the two channels, rather than the entire provided study area, approximately 400 km2. The second goal is to create and analyze accurate LiDAR-based terrains of the Elwha River in order to identify and quantify LWD within. To determine the validity of this method, the results of LWD detection (spatial locations) and quantification (volume) are then compared with Washington field data. Therefore, in chapter 2, I highlight the image processing techniques used to delineate the stream channels, and consequently, the methods applied to identify and quantify in-stream woods. The results are discussed in Chapter 3, and subsequently, the limitations, future work, and conclusion are presented in chapters 4, 5, and 6, respectively. ? ? 14 ?2. Methods 2.1 Image Analysis Defining the stream channel?s boundaries before undertaking the intensive LiDAR processing can confine the analysis to only stream channels, allowing for an efficient use of computing resources. This is a major advantage as opposed to processing an entire study area of, for example, 400 km2 (the entire coverage of HWP LiDAR dataset). HWP provided shapefiles of up-to-date boundary definitions for the Wildhay River and Moberly Creek, but no recent shapefile was delivered for the Elwha River. An automated tool to delineate the stream channel boundaries could be extremely efficient compared to manual digitization. Hence, I focused on finding a unified framework that could accurately segment the boundaries of stream channels despite their geomorphologies and possible canopy-obstruction. Because orthophotos and shapefiles were required for validation, and they were provided only for the Hinton dataset, I applied the methodology only to the Wildhay River and Moberly Creek. 2.1.1 Data Preparation The initial step was to extract the intensity values from the top surface of the LiDAR point cloud, on which further image processing and analysis could be performed. For efficiency purposes, I selected samples from different spatial locations of the Wildhay River. These samples were meant to represent complex branching and mergers of stream channels, and also included LWD within them. The extracted images were noisy and blocky, where individual pixels were apparent and the image was not smooth. Because water absorbs a significant amount of the (IR) LiDAR pulses, corresponding pixels had no brightness values, particularly where water was flowing. Therefore, in order to minimize the impact of missing pixel data, I replaced all missing pixel values with zero, which is the reflectivity value of water. ? ? 15 ?2.1.2 Non-Linear Diffusion Various smoothing techniques were examined and applied to remove noise from the restored images while preserving strong edges, a.k.a clear region boundaries, between water and land pixels. The goal was to have smooth intra-regions of water and land pixels, but with maintaining sharp intensity-value transition in-between. Traditional smoothing techniques smoothed the entire image with no emphasis on region boundaries. For instance, the initial phases of scale-space filtering, as introduced by Witkin (1983), was through convolving an original image with Gaussian kernel of variance t to produce a family with derived images that have coarser resolution (Equation 1.1). As illustrated in the equation, a Gaussian mask (G) with scale level (t) is used to smooth the 2-dimensional image (I0) in order to produce a linearly smoothed image (I). A value of 0 for the scale parameter (t) means the final image (I) is equivalent to the original image (I0), and as the scale parameter increases, the final image becomes smoother and much different than the original image. Gaussian Kernel smoothing: (1.1) Where : Initial image I : Final smoothed image G : Gaussian kernel or mask Perona and Malik (1990) later applied a new algorithm that blurred internal regions while preserving region boundaries and sharp edges (Equation 1.2). The gradient operator ( ) works as an edge estimator that searches for sharp color transition?region boundaries?in the image. The Laplacian operator ( ) sharpens and enhances those found boundaries. Knowing region boundaries in the image, the smoothing within regions occurs based on the diffusion coefficient (c (x, y, t)). The equation does not encourage overall blurring of the image, as the diffusion stops when there is a strong edge and no smoothing occurs across boundaries. I(x, y,t) = I0(x, y)?G(x, y;t)I0?? ? 16 ?Proposed anisotropic Diffusion smoothing: (1.2) Where div: Divergence operator : Laplacian operator : Gradient operator c : diffusion coefficient According to Perona and Malik (1990), the proposed anisotropic diffusion algorithm encourages intra-region smoothing by setting the diffusion coefficient to 0 at region boundaries and to 1 inside each region. For boundary detection at multiple scales, an edge-estimate function based on image intensity difference was considered. The accuracy of the edge-estimate function dictated how the algorithm satisfies criteria such as causality, immediate localization, and piecewise smoothing. The causality criterion ensures that the edge-preserving smoothing does not introduce new features to the final image. Selection of the diffusion coefficient as a function of the image gradient guarantees a stable edge enhancement throughout the smoothing process (for further information regarding the proofs of these properties, see the appendix of Perona and Malik (1990)). I implemented the Perona and Malik algorithm to smooth water and land regions separately while maintaining sharp edges?clear boundaries?in-between. A Matlab implementation of the Perona and Malik algorithm, developed by Guy Gilboa, was applied. The algorithm ran with experimental parameter values 50 and 10 for the iterations and PM (Perona & Malik) threshold, respectively. ? It= div(c(x, y,t)?I ) = c(x, y,t)?I +?c i?I?? ? 17 ?2.1.3 Stream Channel Segmentation Techniques Two approaches were taken to segment the stream channels: (1) define the stream channel from bank-to-bank and (2) consider just the current water flow in the channel. 2.1.3.1 Vessel Enhancement Filter One approach to outline the stream channels is to consider the channel boundaries at a bankfull stage or discharge. Since stream channels are tubular-structures with often complex branching and merging, it was worthwhile considering the application of a vessel enhancement filter technique as a pre-processing step to the actual channel delineation. A state-of-the-art approach is the multi-scale enhancement filter proposed by Frangi et al. (1998) that has a smooth Gaussian filter and the ability to suppress noise and background for 2-D and 3-D images. Frangi et al.?s multi-scale enhancement filter reformulates the vessel enhancement problem to searching the image for and enhancing a predefined geometric structure?tubular in this case. This method uses a geometric interpretation of the eigenvalues of the hessian matrix to probabilistically determine vessel existence. ?1 represents the principal direction along the vessel with minimal intensity changes as shown in Figure 4, whereas ?2 represents the direction towards the edges, the sides of the vessel. For a tubular-structure filter enhancement, ?2 takes on a high value as it describes the gradient change towards side-edges (Table 2). ?1 on the other hand, indicates the gradient change along the vessel, which tends to be low along vessel-like structures. In simple terms, a tubular structure can be defined as a similar and consistent color along one direction and a sharp color transition along the other direction. The signs in Table 2 are related to whether you?re looking for bright or dark edges. ? 18 ? ?Figure 4: Eigen values principle directions. ?1 represents the gradient (color) change along the tubular structure while ?2 represents the gradient change across the tubular structure. ? ??1 ?2 Orientation Pattern Low High - Bright Tubular-structure Low High + Dark Tubular-structure Table 2: Eigenvalues for different options of tubular structures. A dark tubular structure indicates a low change of color along the shape and a high positive value for the change of color across that same shape. According to Frangi et al., the Vesselness2 response for a dark stream channel in a white background in a 2-D image can be described Equation (1.3). According to equation (1.3), there is no enhancement across the boundaries (?2 >0), but there is along the vessel with a magnitude described below. ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ?2 ?Vesselness ?is ?a ?term ?that ?refers ?to ?the ?vessel ?enhancement ?filter. ? ? 19 ? (1.3) Where: is a geometric ratio that accounts for the deviation from a blob-like structure. S = (?1^2+ ?2^2) , is a measure of second order structureness used to minimize the background structure and is based on the Frobenius matrix norm. ? and c are arbitrary constants for enhancement filter correction. Experimentally, I applied the filter with parameter values 5 and 15, respectively. 2.1.3.2 Region-Growing3 An alternative way to identify stream channels is to delineate the stream based on the current water flow. This scenario would be comparable to the previous one during high-flow seasons, where water is actually flowing from bank-to-bank. During the summer and low-flow periods, the distinction between the two definitions becomes apparent. I segmented the water-flow areas in the Wildhay River through combinations of region-growing and morphological operators4. Patterns of morphological operators, such as opening and closing, were used to remove small gaps in channel boundaries. Once the region-growing step was completed, the channel boundaries were delineated accordingly. 2.1.4 Logjam Detection Using Spectral Signature One method to identify locations of in-stream woods is to analyze the spectral signatures of the LiDAR intensity images. Because water has very low reflectance on infrared ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ?3 ?Region-??Growing ?is ?a ?pixel-??based ?image ?segmentation ?technique ?that ?requires ?the ?selection ?of ?a ?seed ?point. ?4 ?Morphological ?operators ?are ?a ?set ?of ?operators ?(e.g., ?erosion, ?dilation, ?opening, ?and ?closing) ?that ?transform ?digital ?images ?on ?the ?basis ?of ?geometrical ?structures. ?V0(s) =0,if?2> 0,exp(??B22? 2 )(1? exp(? S22c2 ))??????B=?1?2,S = ?12+ ?22( ) ? 20 ?images, compared to nearby objects within the channel, LWD could exhibit a unique spectral signature, especially if they are adjacent to dark water pixels. With reference to Hinton orthophotos, a visual examination of the LWD spectral signatures was carried out in an attempt to find a distinct signature for woods within channels. Accordingly, the boundaries of identified LWD were also highlighted for further length and width comparisons with terrain analysis results. ? ?2.2 Terrain Analysis The methods applied differed slightly between study areas, depending on the dataset and resolution available. In Hinton, for instance, I attempted to identify major locations of in-stream woods, but because there was no ground-truthing data, I was not able to validate the results. However, in Elwha River, both identifying and quantifying in-stream LWD was possible. Terrains described in 3-dimensions (XYZ) are obviously more realistic and can provide more information than 2-D terrains (XY). Using customary 2D orthophotos, researchers continue to detect unobstructed LWD using a variety of image classification techniques. However, 3D terrain analysis (using a point cloud) is capable of providing information about vegetation-obstructed LWD and their volume. LiDAR opens up the opportunity to use a 3D analysis tool that can provide different digital elevation models describing various topographies, such as ground, vegetation, buildings, water, etc. However, it has been found that the quality of the LiDAR-derived ground surface directly affects the quality of the other LiDAR-derived height models used in subsequent vegetation modeling (Evans and Hudak, 2007). Therefore, I first attempt to derive an optimum ground surface5 (DTM) and a surface model6 (DSM) that could describe in-stream woods. ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ?5 ?Optimum ?DTM ?is ?a ?derived ?surface ?that ?best ?describes ?the ?ground ?surface ?despite ?the ?existence ?of ?features ?above ?ground, ?such ?as ?trees, ?logjams, ?shrubs, ?etc. ?6 ?Optimum ?DSM ?is ?a ?derived ?surface ?that ?best ?depicts ?the ?top ?surface ?of ?features ?above ?ground, ?such ?as ?logjams, ?low, ?medium, ?or ?high ?vegetation. ? ? 21 ?2.2.1 Producing Digital Terrain Model (DTM) One of the key benefits of using LiDAR is its capability to produce a bare-earth model despite existing forests, buildings, and other features on (and above) the ground. The quality of the bare-earth model, however, depends on the point density of the LiDAR data. Most LiDAR data providers preprocess the data and assign a classification for each LiDAR return based on a standard of point classification (Table 3). Every LiDAR return from the point cloud would be classified as ground, building, or water, etc. Data providers may also deliver a number of DTMs, DSMs, and intensity images in addition to the point cloud. Such deliverables are often general artifacts that suit common user needs such as visualization and preliminary analysis. Nonetheless, for detailed analysis and in order to have more control over the results, one should start with the raw point cloud and process the data as required. ? Classification Value (Bits 0:4) Meaning 0 Created, never classified 1 Unclassified1 2 Ground 3 Low Vegetation 4 Medium Vegetation 5 High Vegetation 6 Building 7 Low Point (noise) 8 Model Key-point (mass point) 9 Water 10 Reserved for ASPRS Definition 11 Reserved for ASPRS Definition 12 Overlap Points 13-31 Reserved for ASPRS Definition Table 3: The Standard LiDAR Point Classification of the American Society for Photogrammetry and Remote Sensing (ASPRS)(Graham, 2012). I tried both options to produce an optimal DTM?building various DTMs from the raw point cloud as well as utilizing the post-processing point classification information provided. The goal was to have the overall smoothest DTM with the lowest elevation ? 22 ?underneath vegetation and logjams. Next are descriptions of the techniques that were used to create candidate DTM?s. The first common approach to construct a DTM is to consider the last LiDAR returns as ground returns. However, due to presence of dense vegetation cover and dense LWD?which LiDAR pulses could not penetrate?this method proved unsatisfactory and created non-smooth and unrealistic ground terrain. Such a common approach may work well for unobstructed stream channels or those not covered by dense vegetation, which was not the case here. Therefore, given the poor quality of the results, this method?s output was disregarded and the focus shifted to the following two methods. As described by Evans and Hudak (2007), multiscale curvature classification (MCC) is a useful technique that allows the user to define parameters in order to produce a meaningful DTM from a raw and non-classified point cloud. It is an iterative algorithm for classifying a complex LiDAR point cloud as ground and non-ground returns based on a positive curvature threshold in forested environments with complex terrain (Evans and Hudak, 2007). For instance, a higher curvature threshold is typically assigned when deriving a rugged terrain, allowing for higher transitions in slopes whereas a smaller curvature value is used in flat terrains. Hence, the user-defined parameters extend the utility of the MCC by allowing for more flexible and topographic-dependent ground surfaces. The MCC algorithm requires minimal user input represented by a LiDAR point cloud, in addition to scale, curvature, and cell size parameters in order to maintain repeatability and flexibility (Table 4). The application of different values for the scale, curvature, and cell size parameters (Table 4) yielded 50 possible DTMs, from which the DTM that best satisfied my requirements in terms of elevation and smoothness was selected. ? ? ? 23 ?Parameter Definition Value Ranges Scale An input parameter that is a function of the scale of objects on the ground (e.g., trees and LWD) as well as the LiDAR point spacing. 1 - 6 Curvature A positive tolerance threshold that if exceeded, a LiDAR return is classified as non-ground return. Curvature threshold value, also, depends on the landscape configuration and vegetation structure. 0.3 - 5 Cell-size Refers to the resolution of the output grid 1 and 3 Table 4: The definitions and values considered for the user-parameters of the MCC algorithm. The third approach was to rely on the post-processed LiDAR point classification defined by the data provider. Both Hinton and Elwha LiDAR datasets provided point classification information (e.g., ground, water, low vegetation, high vegetation etc.; Table 3) and based on which their classifications, corresponding DTMs were created. Then, these DTMs were visually examined thoroughly over stream channels, channel banks, floodplains, forested areas, and LWD. An illustration showing the classified ground returns is presented in Figure 5, where all of the yellow dots represent the ground surface (classified as such by the data provider). By overlaying the ground returns over an orthophoto, the example shows that the majority of ground returns reflected off the water surface, floodplain, gaps between the canopy, and gaps between the LWD in the middle of the image. The results of the two above-mentioned methods were two filtered subsets from the entire LiDAR point cloud that best characterize the ground surface. Subsequently, the actual DTM was generated through converting each filtered subset to a raster surface of a given size, 0.09-m2 for Elwha DTM and 1-m2 for Hinton DTM. Those units were assigned differently based on the point density of each acquired datasets (8.9-27 returns per m2 for Elwha dataset and 1 return per m2 for Hinton dataset). In case a raster unit had more than 1 LiDAR return, I averaged their values to obtain a single mean height. Lastly, I created hillshade reliefs and classified the raster based on height for better LWD visualization and quantification. ? 24 ? Figure 5: From the entire point cloud of the Wildhay River, yellow points are the extracted point-returns that were classified as ground returns. In general, ground returns were located on water surface, floodplain, and in between gaps of canopy and LWD. ?2.2.2 Producing Digital Surface Model (DSM) The DSM, in its simplest form, describes the top surface of features above ground or the terrain as seen from a flying aircraft. There can be variations to this definition depending on the features or surface of interest. The initial DSM in this project was a simple raster surface created based on the first LiDAR point returns. Nonetheless, when compared with actual field data, this DSM failed to identify the exact boundaries of those in-stream woods that are completely or partially obstructed by vegetation covers. Therefore, an attempt was made to produce a DSM that would be better suited for the identification of partial and completely obstructed LWD. To do this, I took advantage of the multiple return nature of the LiDAR dataset. Thus, to overcome the canopy obstruction, I filtered out those LiDAR point returns that would be unlikely to describe LWD, whether it was low, medium, or high vegetation. Hence, only LiDAR returns below a certain height threshold were selected from the entire LiDAR point cloud. Figure 6 is a side-view illustration of the filtration process applied to the point cloud. As illustrated in Figure 6A, the tree canopy partially obstructs the view of the LWD and, ? 25 ?hence, the DSM created based on this non-filtered LiDAR data would miss the complete boundaries of that LWD, as described by the red stroke. A failure to capture the complete boundaries of a LWD would lead to wrong predictions about the length, width, and height of that LWD. However, as illustrated in Figure 6B, by selective filtering, I was able to disregard all LiDAR point returns that were higher than 3 m. As a result, the tree canopy was excluded, because it was higher than the threshold and, subsequently, the constructed DSM captured the complete LWD boundaries, as described by the red stroke in Figure 6B. Figure 6: (A) shows how high vegetation could hinder the detection of the LWD boundaries. (B) illustrates how feasible is the delineation after excluding point-returns higher than 3 m above ground level. In order to identify locations of in-stream woods, I assigned low values to the height threshold, generally 1, 2, and 3 m. After identifying locations of LWD, I aimed to capture the highest surface that could describe every individual LWD. Although this step facilitated the LWD height calculation, as described in the following section, it may also introduce uncertainty to the height measurements. The height threshold could be underestimated if the uppermost shape of the LWD is not clear and distinct, leading to some LWD being cut off. For instance, LiDAR measurements may report a height of 2 m for a logjam, where in fact, it is 3 m in height. Nonetheless, the filtration technique was applied to all identified in-stream wood locations and every individual logjam was assigned a height threshold. Since field data reported a maximum height of 5.5 m across all in-stream woods, the height threshold values ranged between 1-6 m. ? 26 ?2.2.3 Logjam Volume Detection In general, the shape and brightness values7 could be utilized to identify the locations of in-stream woods and their boundaries. As shown in Figure 7A, the shape of logjams is distinct and so was their spectral signature relative to their immediate surroundings (Figure 7B). Based on such information, I estimated the volume of identified in-stream woods through measurement of their dimensions (length, width and height). In an ideal scenario, with no vegetation cover obstructing the LiDAR sensor view, the boundaries of a LWD could be easily and distinctly outlined, as shown in Figure 7A. Given the irregular shape of logjams, there could be variability in measuring the length and width. However, I used the approach followed in the fieldwork (Vivian Leung, Personnel Communication, March 2013), where the length was the distance between a LWD start and end points along the stream channel. Similarly, the width was the distance between the two endpoints across the stream channel. In a few instances, the length and width, based on previous definitions, were interchanged when the distance along the channel was shorter than it across the channel. ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ?7 ?Brightness ?value ?is ?a ?value ?assigned ?to ?each ?pixel ?in ?an ?image ?to ?describe ?its ?reflectance ?characteristics. ?Typically, ?the ?higher ?the ?value, ?with ?respect ?to ?the ?brightness ?range, ?the ?brighter ?the ?pixel, ?and ?vice ?versa. ? ? 27 ? Figure 7: (A) is a hillshade illustration of an ideal scenario of mid-channel LWD. (B) is the intensity image of the same spatial location showing bright pixels for LWD. Figure ?8: ?(A) ?is ?a ?Height ?Model ?showing ?the ?result ?of ?a ?raster ?math ?calculation, ?subtracting ?the ?DTM ?from ?DSM ?model. ?(B) ?is ?a ?reference ?hillshade ?image ?based ?on ?the ?DSM. ? In the same simple scenario, the height of the LWD was obtained through raster surface subtraction operation, so the previously generated optimal DTM was subtracted from the filtered DSM in order to obtain a height model of the logjams within the channels, ? 28 ?(Figure 8). This height model was classified by assigning different colors to different heights. For instance, the LWD in the middle of the channel in Figure 8A was more than 3.4 m in height. The presence of vegetation cover that obstructs the sensor view introduced more complexities to LWD detection and volume estimation. Such complexities varied widely depending on how much of those logjams were obstructed and how dense was the vegetation cover. Therefore, the level of uncertainty rose and in many situations was proportional to the percentage of view-obstruction. Nonetheless, the height of each in-stream logjam was calculated by subtracting the optimal DTM from the corresponding DSM produced after the filtration technique. Measuring the length and width of in-stream logjams was a manual process (Figure 7A), which was time inefficient and subjective. However, it was a cost-effective approach compared to traditional fieldwork and, also, allowed for flexibility through offering alternative ways for dimensions measurements. This was a key benefit, particularly in cases where various measurements could be considered, and in such circumstances, I considered the way whose measurements were closest to the field data. ? ? 29 ?3. Results & Discussion 3.1 Image Analysis Results 3.1.1 Image Restoration And Enhancement The image restoration step introduced great enhancement by minimizing noise and outliers on the sample images of the Wildhay River, as shown in Figure 9. The intensity-scale adjustments followed by an erosion operation and low-pass filtering, also, resulted in clear channel boundaries with minimum noise. This restoration phase produced similar improvement to stream channels? images across various spatial locations of the Wildhay River. The techniques applied on the restoration phase are case and image dependent, as different images could require different set of techniques and parameter values. For instance, a higher resolution image may require less restoration than a lower resolution image, and further modification to the smoothing filter and erosion parameters could, also, lead to better results. Although some images, Figure 9 as an example, could have been further enhanced, the achieved level of restoration was sufficient, enabling me to carry out subsequent tasks. Also, the assumption of replacing missing pixel values with 0 initially seemed unrealistic; however, it proved valid after several experimental attempts. For example, when the missing pixel values were replaced with various non-zero intensity values, the output images came out noisier, particularly, in regions where water was flowing. ? 30 ? Figure 9: (A) is the initial image exhibiting noise and missing data. (B) is the restored image. ?3.1.2 Non-Linear Diffusion: Although restored images showed less noise and outliers, groups of similarly adjacent pixels, particularly in-land, formed tubular shapes of different scales and directions. The non-linear diffusion slightly smoothed and distorted those tubular structures while maintaining well-defined boundaries?sharp edges?between water and land regions. An illustration of the impact of Perona implementation of non-linear diffusion is illustrated in Figure 10A. This type of smoothing followed by Frangi et al.?s enhancement filter was able to recognize parts of the channel where LWD extended across the channel from bank-to-bank?a success that Gaussian smoothing did not accomplish. The final channel delineation after the non-linear diffusion (Figure 10A) was compared to the normal Gaussian smoothing result (Figure 10B). Note, in particular, the differences highlighted inside the yellow circles of the bottom images, exhibiting better channel connectivity and, hence, better delineation using the non-linear diffusion approach. ? 31 ? Figure 10: (A) shows the impact of non-linear diffusion that intuitively described in-stream woods as part of the channel. (B) illustrates the lack of traditional Gaussian smoothing to include such LWD as part of the channel. The non-linear diffusion, while maintaining strong edges, allowed for continuous blurring and elimination of well-defined tubular structures around the stream channel. Such continuous blurring presented the stream channel as the dominant well-defined tubular structure in the image, a key-preprocessing step for the subsequent enhancement filter. The selection of parameters for the non-linear diffusion implementations, such as the number of iterations and PM threshold, was experimental and depended on various factors, as different parts of the Wildhay River were assigned different parameters? values. Also, implementation parameters varied slightly depending on the width of the channel to be detected and, accordingly, smaller values were assigned to creeks and larger values to rivers. ? 32 ?3.1.3 Enhancement Filter Overall, the Frangi et al.?s enhancement filter exhibited significant achievements through considerable suppression of noise and background, as illustrated in Figure 11. However, a few sample images of the Wildhay River showed some artifacts (Figure 12B) despite the use of different values for the filter?s parameters. Figure 12C is an illustration of the channel boundaries produced by the enhancement filter (Figure 12B) overlaid on the original image for validation purposes. In those sample images of the Wildhay River, the enhancement filter provided accurate segmentation of the main stream channel but with non-uniform and scattered artifacts. Despite the overall accurate segmentation exhibited by Frangi et al.?s enhancement filter, the outcomes were clearly inconsistent as the experimental results showed classic segmentation in many samples of the Wildhay River, but in few others, the channel boundaries were somewhat less recognized and enhanced. Also, the enhancement filter accepted various deviations of tubular-structures that were not part of the channel itself. A correction step could be to penalize for such deviations and ensure smoothness and regularity of the enhanced shape. Poon et al. (2007) modeled the Live-Vessel program with internal costs that ensured gradual vessel radius, fluctuations in vessel direction, and smoother medial axis. Such restrictions could contribute to the accuracy of my solution. ? 33 ? Figure 11: The enhancement filter applied to the Wildhay River merging with Moberly Creek. (A) is the original image before the enhancement filter. (B) is the image after the enhancement filter. Figure 12: An illustration of a segment of the Wildhay River channel enhancement filter. (A) is the original image. (B) is the output of enhancement filter showing scattered artifacts around the stream channel. (C) is the channel segmentation overlaid on the original image. ?3.1.4 Region-Growing Due to low-noise and high-connectivity of water pixels, the region-growing technique yielded a consistent segmentation of the Wildhay River (Figure 13). The segmented stream channels described the boundaries of the current water flow rather than the bank-to-bank channel. Therefore, this channel definition excluded sandbars and LWD within the channel and only included connected water pixels. In Figure 13B, the boundaries ? 34 ?produced by the region-growing were overlaid the original image (Figure 13A), demonstrating a detailed segmentation of the water pixels in the Wildhay River. Figure 13: An illustration of a channel segmentation using region-growing technique. (A) is the original image. (B) is the region-growing boundaries results overlaid the original image. ?3.1.5 Logjam Identification Analyzing the spectral signatures of the Wildhay River sample images resulted in different brightness for different features and landscapes. As illustrated in Figure 13A, the darkest pixels of the image mostly represented water and the bars generally exhibited the second darkest parts of the channel. Signatures in the range of 95-128 in the 8-bit gray-scale intensity (0-255) highlighted in-stream woods. Based on such a manual image classification, an experimental association between image features and landscapes and their corresponding pixel brightness range can be developed (Table 5). Nonetheless, the classification scheme presented in Table 5 cannot be generalized beyond this thesis, as different image processing techniques may yield different ranges. Feature / Landscape Brightness Value Floodplain 130 - 158 Healthy vegetation 160 - 191 In-stream woods 95 - 128 Sandbars 30 - 65 Water 0-5 Table 5: Image features and their corresponding experimental intensity values (0-255). ? ? 35 ?Based on the intensity classification presented (Table 5), I was able to identify 9 more non-obstructed, 15 less partially obstructed, and 8 less completely obstructed LWD than reported by the field data (Table 6). Relative to fieldwork, the intensity image approach was the least effective detection technique among the other methods, as it reported 14 less LWD than the total reported in fieldwork, all of which are associated with partial and complete canopy obstruction. Nonetheless, intensity classification is a valuable and maybe more advantageous than fieldwork in non-obstructed environments. ? LiDAR measurements Obstruction/ methods Fieldwork Intensity Images Filtered LiDAR Non-Filtered LiDAR None 17 26 29 29 Partial 35 20 38 38 Complete 8 0 5 0 Total LWD Identified 60 46 72 67 Error of Omission 21.6% 5% 13.3% Error of Commission 15% 25% 25% Overall Improvement Over Fieldwork -6.6% 20% 11.7% Table 6: Number of in-stream woods identified by each of the different methods attempted. The omission and commission errors are based on the fieldwork counts. The overall improvement is the result of subtracting the omission error from the commission error. ?3.2 DTM Comparisons From the 50 possible candidate DTM?s created by MCC in the Elwha River, the lowest and smoothest was obtained with a scale value equal to 2, curvature value equal to 4, and cell size equal to 1. In Table 7, a comparison is made between the DTMs produced by the selected MMC and the point classification information. After preliminary examinations, both DTMs provided similar and consistent results on water and floodplain surfaces and, hence, the points? coordinates in Table 7 refer to various LWD midpoints that are spatially spread along the Elwha River (Figure 14). On those points, it was found that the MCC method generated a relatively higher DTM than the point classification method. On average, the DTM for point classification was 6% lower than the MCC-derived DTM in 95% of all selected locations. One reason for the overestimation of the actual ground level using the MCC method could be an exaggerated curvature threshold. The initial ? 36 ?goal was to obtain the lowest and smoothest possible DTM and, hence, the point classification method clearly yielded the optimal solution, based on which the subsequent analysis was built. ? Point Coordinates DTM Height Comparison (m) Difference (cm) MCC - Classification Point X Y MCC Classification 1 892076.016 1039174.769 4.31 3.69 62 2 892262.927 1039435.158 4.11 3.70 41 3 892342.44 1039483.682 3.61 3.10 52 4 892328.552 1039551.738 3.78 3.55 23 5 892122.996 1039189.238 3.57 3.42 15 6 892748.039 1038703.604 3.40 3.35 5 7 892593.96 1038339.021 5.88 5.06 83 8 892853.291 1037653.474 6.01 5.65 36 9 893343.482 1035683.726 10.35 9.38 98 10 893681.85 1035103.054 9.97 9.89 8 11 894258.239 1033940.902 11.60 11.43 17 12 893414.489 1033156.179 13.81 14.59 -78 13 893954.42 1032894.027 13.61 13.14 47 14 894421.26 1032237.082 15.01 14.72 29 15 894521.607 1032082.568 15.41 15.22 19 16 894497.302 1031641.075 15.56 15.21 34 17 894210.374 1030869.617 17.57 16.95 62 18 894450.062 1029566.179 19.34 18.72 62 19 894523.499 1029295.659 19.37 19.20 17 20 894296.885 1024834.148 27.23 26.55 69 Table 7: Height comparison of selected LWD midpoints between Multicurvature Classification algorithm (MCC) and the point classification. ? ? 37 ? ?Figure 14: The spatially spread LWD-midpoints on which candidate DTMs were compared. The value ranges are based on the absolute difference between the two DTMs in centimeters. ?3.3 DSM Results In Elwha River, LiDAR demonstrated a more comprehensive and multiscale approach compared to fieldwork, as it was able to detect more than 15 new in-stream logjams that were not initially reported in the field survey. Some of those unreported logjams were either small in size or likely difficult to reach. In fact, small-scales in-stream woods?down to single wood pieces?were identified, using the 0.3-m resolution Elwha River dataset, however, only logjams greater than a size threshold (2 m2) were reported as LiDAR measurements. 3.3.1 Factors In LWD Detection The spatial location of in-stream woods had a significant influence on their detection and volume estimation. Those detected logjams varied in spatial settings with reference to the stream channel and to existing vegetation cover. For instance, with respect to the stream ? 38 ?channel orientation, some LWD lied in the middle of the channel while others were pushed towards the channel banks and bars. The vegetation cover obstructed the view of most identified in-stream woods (Table 6), and the percentage of obstruction varied widely. Accordingly, I classified identified in-stream woods as completely, partially, or not covered by vegetation. The volume estimation of completely uncovered LWD was a straightforward task, and their parameters closely matched the field data. However, as the vegetation-obstruction increased, the prediction certainty dropped and, ultimately, few logjams were not identified as the LiDAR pulses could not penetrate dense and closed canopies. Another major factor influencing the accuracy of the LWD detection was the data resolution. Due to the difference in data resolution between Hinton and Elwha datasets, the shape of woods in the Elwha River was more apparent than in the Hinton stream channels. Single woods and the patterns of the accumulating woods were also visible in the Elwha River. Higher point densities led to higher raster surfaces? resolution and provided more details about LWD heights, leading to more accurate dimensional predictions. Nonetheless, the detection of LWD is not a function of the channel width, but rather is a function of the shape and size of logjams. Therefore, a discernible logjam is detectable by LiDAR-based terrain analysis regardless of the size of the stream channel. 3.3.2 Unfiltered LiDAR Returns A comparison was conducted between the field data and the LiDAR results based on two criteria?count of identified in-stream woods and accuracy of their dimensions (length, width, and height). Using the unfiltered LiDAR approach, I was able to detect 52 LWD locations from a total of 60 reported in the field?an omission error of 13.3% (Table 6). However, unfiltered LiDAR was able to detect 15 new logjams that were not reported in the field?a commission error of 25% (Table 6). For every detected LWD location, by LiDAR and fieldwork, dimensional measurements were obtained and compared across the three categories of view-obstruction (Figure 15). ? 39 ?Ideally, one would expect all points to straddle the 1:1 line, indicating that field and LiDAR measurements closely matched. In Figure 15, I present scatter plots comparing the length, width, and height of all LWD measured using fieldwork versus the unfiltered approach. As for the length plot (Figure 15A), in-stream woods with no vegetation cover were tightly scattered around the 1:1 line, indicating that both length measurements were similar. Relative to the field data, LiDAR-derived measurements underestimated partially obstructed LWD and did not report any measurements for completely obstructed logjams. One reason for the underestimation of LiDAR-derived measurements is the canopy obstruction, as the LWD boundaries were not entirely captured, leading to misjudged and undervalued length and width. As for the width, the points were less clustered and more scattered compared to the length?s plot (Figure 15-A). A likely result of canopy obstruction, LiDAR-derived measurements underestimated the width of those partially obstructed LWD reported in the field. Also, no LiDAR-derived measurements were reported for those obstructed LWD and, hence, their corresponding points were laid on the X-axis (Figure 15-A). Predicting height of in-stream woods was more difficult, as a surface profile of such logjams would show wide variations of height values?making it more difficult for the field crew to determine a single height. In general, LWD tend to start low at its edges and continue to rise towards the midpoints. The field crew estimated the height using a LRF while the LiDAR-derived measurements considered the highest point of the LWD. Overall, both measurements were comparable and showed spread around the 1:1 line for non-obstructed LWD (Figure 15A). However, canopy obstruction led to higher height differences in partially and completely obstructed LWD. To summarize, the presence of vegetation cover is a major obstacle for the unfiltered LiDAR approach. The unfiltered LiDAR could not identify any completely obstructed LWD and exhibited less accurate dimensional measurements of partially and completely ? 40 ?obstructed LWD. Due to vegetation obstruction, the actual boundaries of those LWD could not be captured leading to overall underestimation of the length, width, and height measurements. ? 41 ?(A) (B) Figure 15: Comparisons of in-stream woods dimensions across the three types of vegetation obstructions. (A) column compares the field data to LiDAR measurements before the filtering while (B) column compares measurements after the filtering. 0"20"40"60"80"100"120"140"160"180"200"0" 50" 100" 150" 200"Field&Data&(m)&LiDAR&(m)&Non3Filtered&LWD&Length&Par,al"None"Complete"View&Obstruc?on&0"20"40"60"80"100"120"140"160"180"200"0" 50" 100" 150" 200"Field&Data&(m)&LiDAR&(m)&Filtered&LWD&Length&Par,al"None"Complete"View&Obstrucon&0"10"20"30"40"50"60"70"80"0" 20" 40" 60" 80"Field&Data&(m)&LiDAR&(m)&Filtered&LWD&Width&Par.al"None"Complete"View&Obstruc:on&0"1"2"3"4"5"6"0" 1" 2" 3" 4" 5" 6"Field&Data&(m)&LiDAR&(m)&Non3Filtered&LWD&Height&Par,al"None"Complete"View&Obstruc@on&0"1"2"3"4"5"6"0" 1" 2" 3" 4" 5" 6"Field&Data&(m)&LiDAR&(m)&Filtered&LWD&Height&Par,al"None"Complete"View&Obstruc