Geo-Interpretation Model for Land-Cover/Land-Use Classification
Example Application
Study sites and Data
Our study area is located at Hong Kong Island, where is south of Hong Kong Region, in southern china. (Figure 3). In this area, the status of topography is very steep and mountainous. Urban area distributes along the fringe of the island, where mainly is dependent to the filling sea to increase the urban region since the first exploring to this land one hundred and fifty years ago.
Figure 3 Study Area in Hong Kong Administration Map.
The data in this experiment includes Remote sensing data and ancillary geographical data. Land-sat TM imagery flown on 3 March 1996 and SPOT-HRV imagery flown on 2 February 1999 were available for the study. The ancillary geographical data is DEM data and its byproducts including slope and aspect data. The land-use map finished by experts vision interpretation also is used in validation and updating to the land-use by land-cover classification outcome by building land-use knowledge base. Figure is the 3D- display image effect of Hong Kong Island, which is overlaid by TM 5,4,3 band colored with Red, Green and Blue respectively. Figure 5 is SPOT HRV imagery covered on Hong Kong Island.
Figure 4 3D display of HK Island overlaid by TM 5,4,3 with R,G,B color.
Figure 5 SPOT-HRV covered in Hong Kong Island (2 Feb. , 1999)
With basis on the vision interpreting to the remote sensing image in Fig. 4,5, we can find out that in Hong Kong Island the land-cover type is very complicated and anfractuous for its steep topography and sinuous distribution of urban region. By multi-platform remote sensing image data and synthetical analysis methods to investigate the status of land-cover/land-use will provided us punctual, dynamical and fast validation and updating information to the advanced land decision and planning is such an urban region. Based on RSIGIM model in this work, we have designed a hierarchical classification system to Hong Kong Island land-cover/land, which is also divided into three levels corresponding with the structure of RSIGIM. Level one is the preprocessing to the spatial data in order to enhance information or do basis preparing works for next stage. Level two is to do the coarse to subtle classification under the RS data sources from the TM to SPOT with supported by suitable neural computing algorithms. Level three is to validate and update the landuse with supported by geo-knowledge processing system under the classifying outcome in level two.
Preprocessing to the spatial data
With supported by the first level of RSIGIM, preprocessing to the spatial data is necessary for doing the basis processing works in order to be prepare for next higher cognition process or enhance the effect of the vision of RS image. The preprocessing process mainly should include image enhancement, geometric correction, statistical transformation, etc. For instance, in this sample, the main preprocessing works for land-cover classification is done as below.
-
Geometric correction. Because of the spatial data being acquired from multi-platform its objection system and spatial resolution may be different, then before doing classifying the geometric correction work should be done. The first step of correction is to select the control points from image at the distinct sites in image, such as at reflection of the edge of shore, intersection point of roads, peak of hill, building and bare point.
-
Principal Component Analysis (PCA),or K-L transformation . In the multi-spectral RS image, correlative degree often exists between bands of image. We can use minimum bands to represent the most information in whole bands after PCA to the original bands .
- DEM preprocessing. The most important restriction to regional differences of land-cover type is from topographical element like altitude, slope roughness and aspect. In the preprocessing process of land-cover classification system, these topological elements can be derived from only DEM data. By these topological elements being added into the classification system with support by topological knowledge base, the accuracy can be improved with great degree and of uncertainty of spatial data can be reduced.
- Mathematical and Statistical Clustering. Before classifying, using all multi-spatial RS data to do clustering analysis in order to make known the complicated degree of the study area is important and necessary. This process can let us comprehend the probable land-cover distribution by RS data without considering the accuracy. The typical clustering methods include K-means, ISODATA and ART, etc. Fig 6 is the output of clustering by method of scale space based hierarchical clustering algorithm which is based on simulating annealing theory.
Figure 6 Classification Map by Clustering Method
