13 2 An Expert System–Based Image Classification for Identifying Wetland-Relat ed Land Cover Types Xiaobin Cai and Xiaoling Chen 2.1 INTRODUCTION Understanding wetland changes in relation to urban development using remote sens- ing techniques is critical to planning ecosystem management and sustainable regional development. The per-pixel classication methods have been used extensively in the- matic information extraction (Wang et al. 2004). The spectral property of surface objects was utilized in remote sensing classication in these techniques. However, the spectral property information usually is not sufcient to generate accurate clas- sication results. Therefore, other types of information, such as DEM (digital eleva- tion model) and road data, have been used to improve classication accuracy. The intrinsic spatial distribution and relationships that exist in geographical objects enable the adoption of spatial information, such as the texture of an object, which was widely used in the process of classication at the pixel level. On the other hand, to better represent spatial relationships, spatial information at the object level was more useful in identifying the distribution patterns of different classes. This idea directly resulted in the invention of object-oriented classication software. Ecogni- tion, a software package, was recently enhanced by integrating the object-oriented classication techniques with its classication procedures. However, Ecognition’s classication procedure is too complicated (Burnett et al. 2003), especially when selecting segmentation parameters. This paper introduces a simple object-oriented method to improve classication accuracy. 2.2 STUDY AREA AND DATA The study area is located in northern Jiangxi Province, China, with a geographic range of E115° 24ʹ to 117° 43ʹ and N27° 57ʹ to 29° 47ʹ. The area covers Poyang Lake and its surrounding wetlands, which is one of 21 wetlands in China that have been designated as having international importance under the Ramsar Convention. © 2008 by Taylor & Francis Group, LLC 14 Wetland and Water Resource Mo deling and Assessment Lushan National Park, one of UNESCO’s world heritage sites, is also located in the study area. The image selected is the Landsat ETM+ image taken on October 28, 2004. The Shuttle Radar Terrain Mission (SRTM)–edited DEM was used as the comple- mentary data. The SRTM DEM data has the highest resolution among datasets with a near-global (i.e., between S 56° and N 60°) coverage. It has sufciently detailed topographic information and can be derived to t the selected image classication. Figure 2.1 shows the edited SRTM DEM in the study area. As the altitude changes from high to low, a color gradient from blue to red was used in the legend to represent corresponding altitude differences. The central part of the image was dominated by Poyang Lake Plain, shown as blue in the DEM. 2.3 METHODOLOGY AND RESULTS Normally it is quite difcult to acquire satisfactory classication accuracy because of the spectral similarity of different objects. By examining spectral characteris- tics, three groups of objects with spectral similarity were identied: (a) wetland and grassland/forest, (b) bare land and developing urban area, and (c) built-up area and muddy beach. Figure 2.2 illustrates the special overlap between the built-up area and the muddy beach. To improve the classication accuracy, the road data and DEM (the geographic information system [GIS] data) were incorporated in the classication process. The maximum likelihood classication was used to identify initial boundaries of objects, with the water bodies being extracted and masked from the image. In this process, the spatial relationships among the classied objects and the GIS data were refer- enced to develop decision rules in the expert system in order to better distinguish the spectrally mixed objects. FIGURE 2.1 The edited SRTM DEM in the study area. © 2008 by Taylor & Francis Group, LLC An Expert System –Based Image Classification 15 2.3.1 WATER BODY IDENTIFICATION AND SUPERVISED CLASSIFICATION Generally, water bodies can be easily identied using the Normalized Difference Water Index (NDWI) as follows (McFeeters 1996, Gao 1996): NDWI = (GREEN − NIR)/(GREEN + NIR) (2.1) where GREEN is the brightness value of the green band and NIR is the brightness value of the near-infrared band. However, it is still difcult to use the NDWI (Nor- malized Difference Water Index) to separate the surface waters with plants from the plants in “wet” lands. To address this problem, the Normalized Difference Vegeta- tion Index (NDVI), a good indicator of plants, was also calculated. NDVI = (NIR − RED)/(NIR + RED) (2.2) where RED is the brightness value of the red band. Then the difference between NDVI and NDWI was used to detect water bodies more accurately. Supervised classication was then employed to extract other relevant cover informa- tion from the image by masking out the water bodies. Seven land cover types, includ- ing built-up area, wetland (with vegetation), grassland, forest, bare land, muddy beach, and farmland were identied. 2.3.2 EXPERT KNOWLEDGE According to spectral analysis of these land covers, the muddy beach is easily con- fused with the built-up area, while muddy beaches usually distribute along water FIGURE 2.2 The spectral comparison of the built-up urban area with the muddy beach. © 2008 by Taylor & Francis Group, LLC 16 Wetland and Water Resource Modeling and Assessment bodies and built-up areas are close to roads and rails. Thus, the spatial adjacency of these two objects to water bodies, the province-level roads, and railways were mea- sured on the relevant thematic map, and were then incorporated in the expert system. Spectrally, the wetland (with vegetation) is similar to the forest and the grassland. However, wetlands are usually located adjacent to waters or muddy beaches, with their elevations lower than surrounding forests and grasslands. The normal water level of Poyang Lake uctuates between 5.9 m and 22.20 m seasonally and the height of the hygrophyte inhabiting the area ranges from 13 m to 16 m (Tan 2002). Therefore, the neighborhood relationship and the wetland classication unit’s mean elevation were used as criteria in the expert system. As the spectral property of the developing urban area is similar to bare land, it was not classied as an individual land cover type in the initial supervised classication. Usually the developing urban area has more stable relationships with roads and existing urban areas than bare land; this fact was used as the decision rule in the classication to distinguish the two types of classes. Figure 2.3 shows the classied results. In the left picture the red part is the developing areas while the existing urban areas and roads were depicted as blue or green lines. The result accords well with the color features on the original ETM+ image. 2.3.3 OBJECT NEIGHBORHOOD SEARCH AND DEM ANALYSIS To acquire the spatial neighborhood relationship at an object level, a clump procedure was performed on the image with ERDAS Imagine. The results included an object ID image, where the pixel value stands for its object ID code. A search was applied to the reference object images, which included a water object image and a road object image at the pixel level, to produce a new neighborhood relationship image. Then the neighborhood ID image was obtained with an overlay of the object ID image and the neighborhood relationship image, where the LUT (look up table) was achieved. A new LUT, including only the pixels with a value larger than zero, was produced by sieving the LUT using the histogram of the neighborhood ID image. The new LUT was loaded into the object ID image to produce a neighborhood object image. With FIGURE 2.3 The identied developing urban area (in the left image, the lighter area is the developing urban area identied; the right image is the Landsat ETM+ RGB image with Bands 4, 3, and 2). (See color insert after p. 162.) © 2008 by Taylor & Francis Group, LLC An Expert System –B ased Image Classification 17 the expert knowledge analysis described in section 2.3.2, ve object-level search images were created through object neighborhood searching. To obtain the altitude property of each classied object, the overlay analysis was utilized between the clump result of the supervised classication and the edited SRTM DEM data, which generated the mean elevation of each classied object. 2.3.4 CONSTRUCTIONOFTHE EXPERT SYSTEM The expert system was constructed with the ERDAS Imagine’s Knowledge Engi- neering function, where the prospective class was used as a hypothesis, the deter- mination criterion as a rule, and the parameter as a variable. In the experiment, the supervised classication result, the ve object-level search images, and the mean elevation in every object clump were induced as the variables. The criteria were expressed as the following rules: If (muddy-ser = T and (sc = g or sc = fr) and men-elv ≤ 20) nalc = wt; If (muddy-ser = T and sc = u and (rod-ser = T or ral-ser = T or urb-ser = T) and wat-ser = T) nalc = m; If (muddy-ser = T and sc = u and (rod-ser = T or ral-ser = T or urb-ser = T) and wat-ser = T) nalc = fm; If (sc = bl and (rod-ser = T or ral-ser = T or urb-ser = T)) nalc = dv; Else nalc = sc. where muddy-ser denotes the muddy beach object-searching image, rod-ser the pro- vincial road object-searching image, and ral-ser the railway object-searching image. T represents the object that was located in the searching threshold; the nalc means the nal expert system classication result, and sc is the supervised classication result. Grassland is denoted by g, forest fr, wetland wt, muddy beach m, built-up urban areas u, farmland fm, bare land bl and developing urban area dv. A classied result based on the expert system was obtained, and the comparison between the expert system classication result and the supervised classication results are shown in Figures 2.4 and 2.5. The pictures illustrate that some errors generated by the per-pixel classication were corrected by the expert system–based classication. For example, the previously misclassifed built-up areas were corrected to the muddy beaches and the previously missing developing urban areas near Nan- chang City were detected by the new method. 2.4 CONCLUSION The rule-based expert system approach could improve the classication of wetland- related objects that have similar spectral characteristics. The object-level spatial searching method, which was incorporated with supervised classication results and the road and elevation information, proved to be effective for deriving a set of object spatial searching images. © 2008 by Taylor & Francis Group, LLC 18 Wetland and Water Resource M odeling and Assessment ACKNOWLEDGMENTS This study was funded by the 973 Program (2003CB415205) of the Chinese Natural Science Foundation and the Open Fund project (200401006(1) of the Key Lab of Poyang Lake Ecological Environment and Resource Development of Chinese Min- istry of Education housed in Jiangxi Normal University. REFERENCES Burnett, Charles, Kiira Aaviksoo, and Stefan Lang. 2003. An object-based methodology of mapping mires using high resolution imagery. Paper presented at Ecohydrological Pro- cesses in Northern Wetlands Tallinn. Estonia: Tartu University Press. Gao, B. 1996. NDWI—a normalized difference water index for remote sensing of vegetation liquid water from space. Remote Sensing of Environment 58:257–266. McFeeters, S. K. 1996. The use of the Normalized Difference Water Index (NDWI) in the delineation of open water features. International Journal of Remote Sensing 17:1425–1432. Water Built-up Urban Welland Mud Beach Forest Grassland Bare Land Farmland Developing Urban Area ( a )( b ) FIGURE 2.4 The overall comparison between the per-pixel classication (a) and the expert system–based classication (b). (See color insert after p. 162.) FIGURE 2.5 The comparison for the Poyang Lake area with the two classication meth- ods (the left image is the per-pixel classication result, the middle one the original Landsat ETM+ image, and the right the expert system–based classication result). (See color insert after p. 162.) © 2008 by Taylor & Francis Group, LLC An Expert System –Based Image Classification 19 Tan, Qulin. 2002. Study on remote sensing change detection and its application to Poyang international importance wetland. Ph.D. dissertation, Institute of Remote Sensing Applications, Chinese Academy of Sciences, 48–49. Wang, Zivu, Wenxia Wei, Shuhe Zhao, and Xiuwan Chen. 2004. Object-oriented classica- tion and application in land use classication using SPOT-5 PAN imagery. Geoscience and Remote Sensing Symposium, IGARSS ’04 Proceedings, Vol. 5, 3158–3160. © 2008 by Taylor & Francis Group, LLC . spectral similarity were identied: (a) wetland and grassland/forest, (b) bare land and developing urban area, and (c) built-up area and muddy beach. Figure 2. 2 illustrates the special overlap. FIGURE 2. 2 The spectral comparison of the built-up urban area with the muddy beach. © 20 08 by Taylor & Francis Group, LLC 16 Wetland and Water Resource Modeling and Assessment bodies and built-up. image by masking out the water bodies. Seven land cover types, includ- ing built-up area, wetland (with vegetation), grassland, forest, bare land, muddy beach, and farmland were identied. 2. 3.2