Object-Based Land Cover Classification Using High-Posting-Density LiDAR Data

Transcription

1 Object-Based Land Cover Classification Using High-Posting-Density LiDAR Data Jungho Im 1 Environmental Resources and Forest Engineering, State University of New York, College of Environmental Sciences and Forestry, One Forestry Drive, Syracuse, New York John R. Jensen and Michael E. Hodgson Department of Geography, University of South Carolina, Columbia, South Carolina Abstract: This study introduces a method for object-based land cover classification based solely on the analysis of LiDAR-derived information i.e., without the use of conventional optical imagery such as aerial photography or multispectral imagery. The method focuses on the relative information content from height, intensity, and shape of features found in the scene. Eight object-based metrics were used to classify the terrain into land cover information: mean height, standard deviation (STDEV) of height, height homogeneity, height contrast, height entropy, height correlation, mean intensity, and compactness. Using machine-learning decision trees, these metrics yielded land cover classification accuracies > 90%. A sensitivity analysis found that mean intensity was the key metric for differentiating between the grass and road/parking lot classes. Mean height was also a contributing discriminator for distinguishing features with different height information, such as between the building and grass classes. The shape- or texture-based metrics did not significantly improve the land cover classifications. The most important three metrics (i.e., mean height, STDEV height, and mean intensity) were sufficient to achieve classification accuracies > 90%. INTRODUCTION LiDAR (Light Detection And Ranging) is an active optical remote sensing system that generally uses near-infrared laser light to measure the range from the sensor to a target on the Earth surface (Jensen, 2007). LiDAR range measurements can be used to identify the elevation of the target as well as its precise planimetric location. These two measurements are the fundamental building blocks for generating digital elevation models (DEMs). Digital elevation information is a critical component in most geographic information system (GIS) databases used by many agencies such as the United States Geological Survey (USGS) and Federal Emergency Management 1 Corresponding author; imj@esf.edu 209 GIScience & Remote Sensing, 2008, 45, No. 2, p DOI: / Copyright 2008 by Bellwether Publishing, Ltd. All rights reserved.

2 210 IM ET AL. Administration (FEMA). DEMs can be subdivided into: (1) digital surface models (DSMs) containing height information of the top surface of all features in the landscape, including vegetation and buildings,; and (2) digital terrain models (DTMs) containing elevation information solely about the bare Earth (i.e., no above-surface heights) surface (Jensen, 2007). Most LiDAR remote sensing systems that are used for terrestrial topographic mapping use near-infrared laser light from 1040 to 1060 nm, whereas blue-green laser light (at approximately 532 nm) is used for bathymetric mapping due to its water penetration capability (Mikhail et al., 2001; Boland et al., 2004). Because each LiDAR point is already georeferenced, LiDAR data can be directly used in spatial analysis without additional geometric correction (Flood and Gutelius, 1997; Jensen and Im, 2007). Most LiDAR remote sensing systems now provide intensity information as well as multiple returns representing surface heights. The return with the maximum energy of all returned signals for a single pulse is usually recorded as the intensity for that pulse (Baltsavias, 1999). Other factors such as gain setting, bidirectional effects, the angle of incidence, and atmospheric dispersion also influence the intensity values recorded. Leonard (2005) points out that systems with automatic gain control adjust the return signal gain in response to changes in target reflectance. Thus, such variability in intensity values can make it problematic to interpret or model the intensity data. LiDAR data have been recently used in a variety of applications including topographic mapping, forest and vegetation, and land cover classification. Some studies have demonstrated the effectiveness of LiDAR for DEM-related terrain mapping (Hodgson et al., 2005; Toyra and Pietroniro, 2005; Raber et al., 2007). Many researchers have adopted LiDAR remote sensing for identifying vegetation structure and tree species (Hill and Thomson, 2005; Suarez et al., 2005; Bork and Su, 2007; Brandtberg, 2007). LiDAR data have also been used for land cover classification (Song et al., 2002; Hodgson et al., 2003; Lee and Shan, 2003; Rottensteiner et al., 2005). Most previous research focused on integrating LiDAR data with other GIS-based ancillary data and/or remote sensing data such as multispectral imagery. LiDAR data have been generally used as ancillary data to improve land cover classification accuracy. Data fusion between LiDAR and other remote sensing data was a critical part of the studies. Some studies fused the data at a pixel level, whereas others performed data fusion at a decision level. LiDAR data obtained at very high posting densities (i.e., < 0.5 m) increases its potential in a variety of applications. High-posting-density LiDAR-derived information may be used to identify not only elevation of a landform but also precise height and shape information of a target on the Earth s surface such as vehicle or tree. The question then becomes, Can such high-posting-density LiDAR data be used as the sole information source in applications such as image classification without other ancillary data? This study explores the applicability of high-posting-density LiDAR data for land cover classifications with an object-based approach focusing on height information. For example, different land cover classes such as buildings and trees may have similar heights yet different shapes. Various height-related metrics such as mean, standard deviation, and textures were extracted for objects as well as a shape index, such as compactness. LiDAR-derived intensity data were also incorporated in the analysis to determine its relative usefulness in land cover classification. In

3 OBJECT-BASED LAND COVER CLASSIFICATION 211 Table 1. Summary of the Nominal LiDAR Data Collection Parameters Nominal LiDAR data collection parameters Date November 14, 2004 Average altitude AGL 700 m Air speed 130 knots Flightline sidelap 70% Nominal posting spacing 0.4 m summary, the objectives of this study are to: (1) explore the applicability of highposting-density LiDAR-derived features for land cover classification without incorporating other remote sensing data, such as multispectral imagery; (2) evaluate the capability of object-based metrics with machine-learning decision trees to classify the LiDAR data; and (3) assess the sensitivity of each object-based metric for land cover classification. Study Area METHODOLOGY The study area was located on the Savannah River National Laboratory (SRNL), a U.S. Department of Energy facility located near Aiken, SC. SRNL has: more than 450 waste sites that store a variety of waste materials; a number of man-made structures such as buildings and roads; water bodies; and forest (Mackey, 1998). Three study sites were selected to demonstrate the capability of high-posting-density LiDAR data for land cover classification. Figure 1 shows the study sites on an orthophoto ( m spatial resolution) collected in 2001 along with the first return LiDAR surface of the sites. LiDAR Data Collection LiDAR data were collected by Sanborn, LLC on November 14, 2004 using an Optech ALTM 2050 LiDAR sensor mounted on a Cessna 337 aircraft. The Optech ALTM 2050 LiDAR sensor collected small footprint first and last returns (x, y, and z) and intensity data using a 1064 nm laser with a pulse repetition frequency (PRF) of 50 khz (Garcia-Quijano et al., 2007). Specifications for the high-posting-density LiDAR data collection are summarized in Table 1. Although the nominal posting spacing was 0.4 m, the actual posting spacing was much finer due to the 70% overlap between multiple flightlines, resulting in an observed posting density of 15.3 points/m 2. The LiDAR data were post-processed to yield x, y, and z coordinates for all first and last returns. Last return data were processed to generate a bare Earth dataset eliminating obstructions on the ground using TerraModel s TerraScan morphological filtering software. The vertical accuracy of the LiDAR data was assessed to be 6 cm RMSE (Garcia-Quijano et al., 2007). A triangular irregular network (TIN) model was

4 212 IM ET AL. Fig. 1. A. Three study sites of the Savannah River National Laboratory (SRNL). B. First return LiDAR surface of Site A. C. First return LiDAR surface of Site B. D. First return LiDAR surface of Site C. used to create the separate DTMs based on either first returns, last returns, bare Earth, or intensity point data. Four DSMs were then created from the TIN models at a spatial resolution of 25 cm. A local height surface was also created by subtracting the bare Earth surface from the first return surface. Figure 2 depicts the flow diagram of the methodology used in the study. A total of five surfaces (i.e., first returns, last returns, bare Earth, height, and intensity) were subsequently used to identify image objects through the image segmentation process. Eight object-based metrics were extracted for use in a machine-learning decision tree classification. Sensitivity analyses were conducted to determine which object-based metrics contributed most to the discrimination of land cover classes. Details are described in the following sections.

5 OBJECT-BASED LAND COVER CLASSIFICATION 213 Fig. 2. Flow diagram of the methodology used in the study. Ground Reference Data Two hundred ground reference points for each study site were randomly generated for the accuracy assessment using a sampling tool developed using Visual Basic in the ArcGIS 9.x environment (Im, 2006). The land cover for each reference point was identified through visual interpretation of a high-spatial-resolution aerial orthophoto and personal communication with SRNL on-site experts. The interpreted land cover was assigned to one of five land classes: (1) building; (2) tree; (3) grass; (4) road/parking lot; and (5) other artificial object. There were no trees in Site C. The other artificial object class included vehicles, small storage containers (e.g., drums), and pipelines. Image Segmentation Definiens (former ecognition) software was used to segment image objects based on the five surface layers generated from the LiDAR returns: first returns, last returns, bare Earth, height, and intensity (Fig. 3). The segmentation process was a bottom-up, region-merging approach where smaller objects were merged into larger

6 214 IM ET AL. Fig. 3. Five layers used to generate image objects during the image segmentation process for Site A.

7 OBJECT-BASED LAND COVER CLASSIFICATION 215 objects based on three parameters (Baatz et al., 2004). A scale parameter determined the maximum allowable change of heterogeneity when several objects were merged (Benz et al., 2004). The segmentation process was terminated when the growth of an object exceeded a scale parameter value. Greater values for the scale parameter resulted in larger objects. Five different scale parameters (i.e., 5, 10, 15, 20, and 25) were tested. A scale parameter of 10 was determined to be the best based on visual interpretation of the resultant image segments. Two more parameters were required to perform image segmentation: color/shape and smoothness/compactness. The color/shape parameter controlled the composition of color versus shape homogeneity during image segmentation. The color information represents elevation at three levels, height, and intensity. The best parameter value for color was 90% and shape was 10%. The smoothness/compactness parameter can be determined when the shape criterion is greater than 0. Several different combinations of the parameters were evaluated based on visual inspection of the resultant objects. A shape criterion of 10% with subdivided smoothness (8%) and compactness (2%) values was selected. Finally, an equal weight (1.0) was assigned to each of the input layers except intensity. The intensity of a laser return was occasionally adjusted manually when collecting the data, which resulted in inconsistent intensity data along the different flight lines. Nevertheless, the intensity surface was used as an input image for image segmentation because it was useful for distinguishing different features with similar heights (e.g., road vs. grass). Thus, a weight of 0.1 was assigned to the intensity surface layer. Object-Based Landscape Ecology Metrics Eight landscape ecology-based metrics were used to classify the image objects created from the LiDAR-derived surface layers. The eight metrics were: (1) mean height; (2) standard deviation (STDEV) height; the texture measures of (3) height homogeneity, (4) height contrast, (5) height entropy, and (6) height correlation; (7) mean intensity; and (8) compactness. Note that the first six metrics were extracted from the height information. The equations for these metrics are presented in Table 2. A suite of very useful texture measures was originally developed by Haralick and associates (Haralick, 1986). The higher-order set of texture measures is based on brightness value spatial-dependency grey-level co-occurrence matrices (GLCM). GLCM is a tabulation of how often different combinations of pixel grey levels occur in a scene (Baatz et al., 2004). The GLCM-derived texture transformations have been widely adopted by the remote sensing community and are often used as an additional feature in multispectral classifications (Franklin et al., 2001; Maillard, 2003). The four textures from the height information (i.e., height homogeneity, height contrast, height entropy, and height correlation) were based on the GLCM-derived texture transformations. Decision Tree Classifications Machine-learning decision trees have been used in remote sensing applications, especially focusing on image classification during the past decade (e.g., Huang and

8 216 IM ET AL. Table 2. Eight Object-Based Metrics Used in the Classification Information category Metric Equation Description Height Mean height STDEV height Height homogeneity i = 0 σ H j = 0 = n H i i μ = 1 H = n i = n ( H i μ H ) ( i j) 2 V ij, ij, = 0 V ij, H i is height of the i th pixel in an object and n is the total number of the pixels in the object. STDEV height indicates variation of height in an object. i is the row number and j is the column number. V i,j is height in the cell i,j of the matrix. n is the number of rows or columns. Homogeneity weights the values decreasing exponentially according to their distance to the diagonal. Height contrast i = 0 j = 0 fi ( j) 2 V ij, ij, = 0 V ij, Contrast is the opposite of homogeneity. It is a measure of the amount of local height variation in the image. It increases exponentially as (i-j) increases. Height entropy i = 0 V ij, V ij, 1n V ij, j = 0 ij, = 0 V ij, ij, = 0 Height entropy is high if the elements of GLCM are distributed equally. It is low if the elements are close to either 0 or 1. Since ln(0) is undefined, it is assumed that 0 ln(0) = 0.

9 OBJECT-BASED LAND COVER CLASSIFICATION 217 Height correlation Intensity Mean intensity i = 0 j = 0 ( i μ )( j μ ) 2 V ij, V ij, ij = σ 2 n I i i μ = 1 1 = n μ = i Height correlation measures the linear dependency of heights of neighboring pixels in an object. I i is intensity of the i th pixel in an object and n is the total number of the pixels in the object. Shape Compactness The compactness c is calculated by the product of the length l and the width c = l m n i = 0 j = 0 V ij, = i μ ij, = 0 V ij,,σ 2 i = 0 j = 0 ( ) 2 m of the corresponding object and divided by the number of its inner pixels n V ij, ij, = 0 V ij, Sources: Baatz et al., 2004; Jensen, 2005.,

10 218 IM ET AL. Jensen, 1997; Hodgson et al., 2003; Im and Jensen, 2005; Im et al., 2008). The popularity of decision trees is largely due to their simplicity and speed for modeling or classifying phenomena under investigation based on example data. Classification literature also points out that one of the advantages of decision trees over traditional statistical techniques is the distribution-free assumptions and independency of features (Quinlan, 2003; Jensen, 2005) Quinlan s C5.0 (RuleQuest Research, 2005) inductive machine-learning decision tree, widely used in remote sensing applications, was used to classify the image objects. Details about C5.0 are found in Quinlan (2003) and Jensen (2005). Im et al. (2008) adopted C5.0 in an object-based analysis to classify bi-temporal imagery. The C5.0 classification was found to yield accuracies as high as the nearest neighbor classifications built in Definiens (former ecognition) software. An inference engine tool was developed to apply decision trees generated from the C5.0 machine-learning to corresponding imagery (Im et al., 2008). Object-Based Metrics RESULTS AND DISCUSSION The eight object-based metric images for Site A are shown in Figure 4. Some land cover classes are easily distinguishable in certain metrics (e.g., mean height and mean intensity) based on visual inspection. The 200 reference points were used to compare mean values of the object-based metrics by land cover class (Table 3; Fig. 5). Summary descriptive findings from the analyses are: Mean Height. Samples of the building and tree classes have high mean height values, whereas the road/parking lot and grass samples yield very low values. This was not surprising. The other artificial object samples ranged from approximately 2 to 3 m of the mean height. It was not easy to distinguish between the building and tree classes, and between the grass and road/parking lot classes solely using the mean height information. STDEV Height. The STDEV height values were generally low in the building, grass, road/parking lot classes while relatively high in the tree and other artificial object categories. STDEV height can easily distinguish between the building and tree samples because the tree samples have much higher STDEV height values than the building samples. However, there was not a substantial difference between the grass and road/parking lot classes. The STDEV height values of the other artificial object samples for Site A were relatively higher than those for Sites B and C because more artificial features different in size and shape (e.g., pipelines, small storage, power lines) were found in Site A. Height Homogeneity. Reference data from the building and other artificial object classes indicated high height homogeneity. The tree and grass samples yielded low height homogeneity. Interestingly, the road/parking lot samples also resulted in low height homogeneity. This may be due to the height errors inherent from the first return and the bare Earth. Height Contrast. The building samples yielded relatively low height contrast for all of the three sites. The tree samples produced low height contrast for Site A but

11 Fig. 4. Spatial distribution of the eight object-based metrics for Site A. OBJECT-BASED LAND COVER CLASSIFICATION 219

12 220 IM ET AL. Table 3. Mean Statistics of the Reference Samples for the Eight Object-Based Metrics Object-based metric a Mean height STDEV height Height homogeneity Height contrast Height entropy Height correlation Mean intensity Compactness Site A Building Tree Grass Road/parking lot Other artificial object Site B Building Tree Grass Road/parking lot Other artificial object Site C Building Grass Road/parking lot Other artificial object a Normality of each metric-class sample was evaluated based on a Kolmogorov-Smirnov test; 57% of the cases indicated that the observed distribution was not normally distributed using a 95% confidence level. Thus, a t-test between each land cover for each metric was not performed. This also justified the use of a machine-learning decision tree algorithm.

13 OBJECT-BASED LAND COVER CLASSIFICATION 221 Fig. 5. Mean statistics distribution of the land cover classes for each object-based metric based on the 200 reference samples for each study site.

14 222 IM ET AL. high height contrast for Sites B and C. There was no consistent pattern in the height contrast for the land cover classes based on visual inspection. Height Entropy. The other artificial object samples resulted in relatively low height entropy. The grass and road/parking lot samples yielded high height entropy. The building samples resulted in high height entropy for Sites A and C. However, the samples resulted in relatively low height entropy for Site B. Height Correlation. The building samples resulted in high height correlation for the three study sites. The other artificial object for Sites B and C yielded higher average height correlations than the samples for Site A because the object features (i.e., vehicle or storage) for Sites B and C are relatively uniform in shape and size. The grass and road/parking lot samples resulted in similar height correlation values for the three sites. Mean Intensity. The mean intensity values were consistent for the land cover categories in the three sites. The building samples yielded the highest mean intensity while the tree and road/parking lot classes resulted in very low mean intensity. The mean intensity was able to distinguish between the grass and road/parking lot samples, where both had very low mean height information. Compactness. The compactness metric did not yield consistent patterns for the three study sites. The building samples resulted in high compactness values only for Site B. The tree samples yielded the lowest compactness for Site A, while the compactness of the tree for Site B was the second highest. The other artificial object class samples produced relatively high compactness for Sites A and C but the lowest compactness for Site B. The grass and road/parking lot samples produced similar compactness values for the three sites. Decision Tree Classifications A total of 316, 172, and 208 training samples were used to generate decision trees for the three study sites, respectively. The training was successful for all of the three sites (<1%). Accuracy assessment of the decision tree classifications for the three sites is presented in Table 4. The decision tree classification for Site A showed good performance, with an overall accuracy of 92.5% and a Kappa coefficient of The building and tree categories were well classified, with producer s and user s accuracies over 95%. Some of the other artificial objects were confused with the grass and road/parking lot objects, yielding lower producer s accuracy (71.9%). There was also some confusion between the grass and road/parking lot classes. The classification for Site B produced an overall accuracy of 94% and a Kappa of There was some confusion between the tree and grass and the tree and other artificial object classes. Like Site A, the other artificial object class yielded relatively lower accuracies. An overall accuracy of 92.5% and a Kappa were achieved for the classification of Site C. Unlike Sites A and B, the other artificial object category was well classified because the class was relatively uniform in shape and size (i.e., trucks, drum-type storage) and there were no trees on the site. Classified images using decision trees for the three study sites are shown in Figure 6. Based on visual inspection, most of the classification errors were found along the boundaries between objects, such as the boundary between a building and road. For example, the height of a building is 10 m and a road next to the building is 0 m.

15 OBJECT-BASED LAND COVER CLASSIFICATION 223 Table 4. Accuracy Assessment Results of the Decision Tree Classifications for the Three Study Sites Site Accuracy statistic a Land cover class Building Tree Grass Road/ parking lot Other artificial object Site A PA (%) UA (%) OA (%) 92.5 Kappa Site B PA (%) UA (%) OA (%) 94 Kappa Site C PA (%) 100 n.a UA (%) 100 n.a OA (%) 92.5 Kappa a PA = producer s accuracy; UA = user s accuracy; OA = overall accuracy. The mixed pixel in the elevation surface may result in a boundary object with a 5 m height, which is possibly confused with another feature such as a tree or a storage tank. Some classification errors were found in roads and parking lots due to heavy dust on the asphalt; asphalt areas with heavy dust were misclassified as grass. Sensitivity Analysis In order to identify the importance of the individual metrics in the classification process, a sensitivity analysis was performed. One metric was excluded and the remaining seven metrics were used for classification, followed by an accuracy assessment. Another metric was excluded and the remaining six metrics plus the previously excluded metric were used to classify the image objects. This process was repeatedly conducted for all the possible cases for the three study sites. The accuracy assessment results are presented in Table 5. For all three sites, the mean intensity was the most important metric in the classification. Mean intensity was the key metric for distinguishing between the grass and road/parking lot classes, which had similar height values. Without mean intensity, the classification accuracies were about 70% and most classification errors were due to confusion between the grass and road/parking lot classes. The second most important metric was mean height. Mean height was useful in distinguishing features with different height information, such as between the building or tree and the grass or road/ parking lot classes.

16 224 IM ET AL. Fig. 6. Decision tree classification results for the three study sites. There was no apparent contribution from the other metrics based on the sensitivity analysis. However, the STDEV height metric was frequently found in all of the decision trees, which indicated that it also contributed to the classification to some extent. Another experiment was performed to identify whether the shape- or texturebased metrics could contribute to improve classification. Only three metrics, i.e., mean height, STDEV height, and mean intensity, were used as input variables in the decision tree classification for the three study sites. The three classifications resulted in as good performance as the classifications using all the metrics (an overall accuracy of 91.5% and a Kappa of for Site A; an overall accuracy of 94.5% and a Kappa of for Site B; an overall accuracy of 92.0% and a Kappa of for Site C). This suggests that the shape- or texture-based metrics may not be necessary to improve classification. Two reasons might help explain the result: (1) only five land cover classes existed in the study sites; and (2) each land cover class had relatively uniform size and shape, so it was not difficult to achieve high classification accuracy

17 OBJECT-BASED LAND COVER CLASSIFICATION 225 Table 5. Sensitivity Analysis to Identify Contribution of Each Object-Based Metric to the Land Cover Classification Site Exclusion Overall accuracy Kappa Site A Mean height STDEV height Height homogeneity Height contrast Height entropy Height correlation Mean intensity Compactness Site B Mean height STDEV height Height homogeneity Height contrast Height entropy Height correlation Mean intensity Compactness Site C Mean height STDEV height Height homogeneity Height contrast Height entropy Height correlation Mean intensity Compactness solely using a few types of LiDAR-derived data. The simple object-based metrics (i.e., mean height, STDEV height, and mean intensity) were sufficient for the classification. However, the classified images derived solely using the three basic metrics were somewhat different from the classified images using the eight metrics along the boundaries based on the visual inspection. The accuracy assessment might not fully identify such a difference due to the lack of the reference data on the boundaries. Shape- or texture-based metrics may be more useful for classification of more diverse landscapes. SUMMARY AND CONCLUSION The land cover classification experiment used only LiDAR-derived metrics obtained from high-posting-density LiDAR data. The key findings are as follows:

18 226 IM ET AL. High-posting (< 0.5 m) density LiDAR data has potential for land cover classification without incorporating other remote sensing data such as high-spatialresolution multispectral imagery. A series of classifications in this study produced high overall accuracies > 90%. Object-oriented analysis using a variety of metrics associated with height, intensity, and shape was successful, resulting in high classification accuracy. The sensitivity analysis identified that mean intensity was the most important metric for the land cover classification. Mean intensity was the key to differentiating between grass and road/parking lot features. Mean height was the second most important metric. Every land cover class had a unique mean height, except the grass and road/parking lot classes. Texture-based metrics did not contribute to improve the classification accuracy. Simple metrics such as mean height, STDEV height, and mean intensity were sufficient for land cover classification in the study. Texture-based metrics might be more useful to classify more complex landscapes. Intensity data were critical to the land cover classification. However, it is noted that they were not consistently recorded over the region. Noise due to inconsistent data collection can be easily found in Figure 3E. More reliable intensity data collection may improve land cover classification. Remote sensing technology continues to evolve. LiDAR data collected in conjunction with metric camera multispectral data is now a reality. Such integration will reduce locational and temporal discrepancy problems associated with fusing different sensor data (e.g., LiDAR + multispectral). Future research will include applications of the method to other environments such as vegetation species identification. REFERENCES Baatz, M., Benz, U., Dehghani, S., Heynen, M., Höltje, A., Hofmann, P., Lingenfelder, I., Mimler, M., Sohlbach, M., Weber, M., and G. Willhauck, 2004, ecognition User Guide 4, Munich, Germany: Definiens Imaging, Germany. Baltsavias, E. P., 1999, Airborne Laser Scanning: Basic Relations and Formulas, ISPRS Journal of Photogrammetry & Remote Sensing, 54: Benz, U., Hofmann, P., Willhauck, G., Lingenfelder, I., and M. Heynen, 2004, Multiresolution, Object-Oriented Fuzzy Analysis of Remote Sensing Data for GIS- Ready Information, ISPRS Journal of Photogrammetry and Remote Sensing, 58: Boland, J. and 16 co-authors, 2004, Cameras and Sensing Systems, in Manual of Photogrammetry, 5 th ed., McGlone J. C. (Ed.), Bethesda, MD: ASPRS, Bork, E. W. and J. G. Su, 2007, Integrating LIDAR Data and Multispectral Imagery for Enhanced Classification of Rangeland Vegetation: A Meta Analysis, Remote Sensing of Environment, in press. Brandtberg, T., 2007, Classifying Individual Tree Species Under Leaf-off and Leafon Conditions Using Airborne LiDAR, ISPRS Journal of Photogrammetry & Remote Sensing, 61:

19 OBJECT-BASED LAND COVER CLASSIFICATION 227 Flood, M. and B. Gutelius, 1997, Commercial Implications of Topographic Terrain Mapping Using Scanning Airborne Laser Radar, Photogrammetric Engineering & Remote Sensing, 63:327. Franklin, S. E., Maudie, A. J., and M. B. Lavigne, 2001, Using Spatial Co-occurrence Texture to Increase Forest Structure and Species Composition Classification Accuracy, Photogrammetric Engineering & Remote Sensing, 67(7): Garcia-Quijano, M. J., Jensen, J. R., Hodgson, M. E., Hadley, B. C., Gladden, J. B., and L. A. Lapine, 2007, Significance of Altitude and Posting-Density on LiDAR-Derived Elevation Accuracy on Hazardous Waste Sites, Photogrammetric Engineering & Remote Sensing, in press. Haralick, R. M., 1986, Statistical Image Texture Analysis, Handbook of Pattern Recognition and Image Processing, Young, T. Y. and K. S. Fu (Eds.), New York, NY: Academic Press. Hill, R. A. and A. G. Thomson, 2005, Mapping Woodland Species Composition and Structure Using Airborne Spectral and LIDAR Data, International Journal of Remote Sensing, 26: Hodgson, M. E., Jensen, J. R., Raber, G., Tullis, J., Davis, B., Thompson, G., and K. Schuckman, 2005, An Evaluation of LIDAR-Derived Elevation and Terrain Slope in Leaf-off Conditions, Photogrammetric Engineering & Remote Sensing, 71(7): Hodgson, M. E., Jensen, J. R., Tullis, J. A., Riordan, K., and R. Archer, 2003, Synergistic Use of LIDAR and Color Aerial Photography for Mapping Urban Parcel Imperviousness, Photogrammetric Engineering & Remote Sensing, 69(9): Huang, X. and J. R. Jensen, 1997, A Machine-Learning Approach to Automated Knowledge-Base Building for Remote Sensing Image Analysis with GIS Data, Photogrammetric Engineering & Remote Sensing, 63: Im, J., 2006, A Remote Sensing Change Detection System Based on Neighborhood/ Object Correlation Image Analysis, Expert Systems, and an Automated Calibration Model, Ph.D. Dissertation, Department of Geography, University of South Carolina. Im, J. and J. R. Jensen, 2005, A Change Detection Model Based on Neighborhood Correlation Image Analysis and Decision Tree Classification, Remote Sensing of Environment, 99: Im, J., Jensen, J. R., and J. A. Tullis, 2008, Object-Based Change Detection Using Correlation Image Analysis and Image Segmentation, International Journal of Remote Sensing, 29(2): Jensen, J. R., 2005, Introductory Digital Image Processing A Remote Sensing Perspective, Upper Saddle River, NJ: Prentice Hall, 526 p. Jensen, J. R., 2007, Remote Sensing of the Environment: An Earth Resource Perspective, Upper Saddle River, NJ: Pearson Prentice-Hall, 592 p. Jensen, J. R. and J. Im, Remote Sensing Change Detection in Urban Environments, in Geo-spatial Technologies in Urban Environments: Policy, Practice, and Pixels, second ed., Jensen, R. R., Gatrell, J. D., and D. McLean (Eds.), Berlin, Germany: Springer-Verlag. pp

20 228 IM ET AL. Lee, D. S. and J. Shan, 2003, Combining LIDAR Elevation Data and IKONOS Multispectral Imagery for Coastal Classification Mapping, Marine Geodesy, 26: Leonard, J., 2005, Technical Approach for LIDAR Acquisition and Processing, Frederick, MD: EarthData Inc., 20 p. Mackey, H. E., 1998, Roles of Historical Photography in Waste Site Characterization, Closure, and Remediation, Aiken, SC: Westinghouse Savannah River Company, WSRC-MS , 13 p. Maillard, P., 2003, Comparing Texture Analysis Methods through Classification, Photogrammetric Engineering & Remote Sensing, 69(4): Mikhail, E. M., Bethel, J. S., and J. C. McGlone, 2001, Introduction to Modern Photogrammetry, New York, NY, John Wiley, 479 p. Quinlan, J. R., 2003, Data Mining Tools See5 and C5.0, St. Ives NSW, Australia: RuleQuest Research, [available online at: info.html, accessed July 10, 2007]. Raber, G., Jensen, J. R., Hodgson, M. E., Tullis, J. A., Davis, B. A., and J. Berglund, 2007, Impact of LIDAR Nominal Posting Density on DEM Accuracy and Flood Zone Delineation, Photogrammetric Engineering & Remote Sensing, 73(7): Rottensteiner, F., Trinder, J., Clode, S., and K. Kubik, 2005, Using the Dempster- Shafer Method for the Fusion of LIDAR Data and Multi-spectral Images for Building Detection, Information Fusion, 6: RuleQuest Research, 2005, C5.0 Release 2.01 (data mining software), St. Ives, Australia: RuleQuest Research. Song, J. H., Han, S. H., Yu, K., and Y. I. Kim, 2002, Assessing the Possibility of Land-Cover Classification Using LIDAR Intensity Data, paper presented at ISPRS Commission III, Symposium, 9 13 September 2002, Graz, Austria. Suarez, J. C., Ontiveros, C., Smith, S., and S. Snape, 2005, Use of Airborne LiDAR and Aerial Photography in the Estimation of Individual Tree Heights in Forestry, Computers & Geosciences, 31(2): Toyra, J. and A. Pietroniro, 2005, Toward Operational Monitoring of a Northern Wetland Using Geomatics-based Techniques, Remote Sensing of Environment, 97: