GISdevelopment.net ---> AARS ---> ACRS 1990 ---> Geology Disaster

Drainage pattern classification by texture analysis

Mitsuharu Tokunaga
Central Computer Services Co. Ltd.
4-3-13 Tranomon, Minatoku, Tokyo 105 Japan

Toshiaki Hashimoto, Shunji Murai
Institute of Industrial Science, University of Tokyo
7-22 Roppongi, Minako, Tokyo 106 Japan


Abstract
Although a geological map is essential to the exploration of underground resources, it is not prepared in some countries, especially in developing classification enables to give effective information for the exploration. The drainage pattern classification has been carried out by an expert fro aerial photographs or satellite images. The process is very time consuming and the results depend on the skillfulness and experience of the interpreter. So it is very desirable to carry out the drainage pattern classification automatically and objectively with a computer. This paper describes the automated drainage pattern classification by texture analysis of DEM from SPOT images.

Method of Texture
The flow of texture analysis is shown in Fig. 1
  1. Generation of DEM

    The DEM on 40m grid size generated by stereo matching using SPOT images. The located of test sties are as follows.

    KITAMI area about 20km * 20km (fig,. 2)
    ASO area about 20km * 20km (fig 3)



    Fig. 1 Flow of Texture Analysis



    With the current program, there can be seen some matching errors.

  2. Grid size of DEM

    When the grid size is small, fine textures (high frequency component) are analysed. Otherwise, rough textures (low frequency component) are analysed. The four case in 40m, 80m, 160m and 320m were selected in this study.

  3. Quantization level of DEM

    The size od co-occurrence matrix is decided on the quantization level of DEM. When the DEM is distributed between 1 and n, the co-occurrence matrix become n*n matrix size, the analyses were carried out with the images level sliced by 10m,20m, 40m and 80m. The size of co-occurrence matrix in eh study area is shown in table 1

    Table 1 Quantitative level
    Level slice pitch Quantization level
    10m 80*80 - 120*120
    20m 40*40 - 60*60
    40m 20*20 - 30*30
    80m 10*10 - 15*15

  4. Window size

    The window size corresponds to the size of calculated area in co-occurrence matrix, therefore the change of window size influence the texture of geographical features. The analyses were carried out on three cases of window size of 5*5, 11*11 and 15*15.

  5. Calculation of co-occurrence Matrix

    The co-occurrence probability of (k,l) in the window is defined as P (k,1). Where k, I : value of DEM. When the direction of co-occurrence is I ad quantization level of DEM is n, Co-occurrence probability Pi (k, l) in the direction I is expressed by the following matrix.




  6. Types of Texture

    The following textures were derived from Co-occurrence matrix

    1. Energy



    2. Entropy



    3. Correlation



    4. Homogeneity



    5. Moment of Inertia




Fig. 3 DEM of ASO Area from SPOT Stereo Images


Fig. 4 Grid size of DEM

Results of Case Studies
  1. Fig 5 shows the change of energy by window size in kitami area. As the window size become larger, the high frequency components become invisible, and classification of drainage systems get better because the noises decrease proportional to the window size.

  2. Fig. 6 shows the change of energy by grid size in KITAMI area. As the grid size become larger, the high frequency components also become invisible. The classification results of 80m grid size was better than the others.

  3. Fig 7 shows the change of energy by quantization level. As the quantization level become smaller, the variation of result become greater. The classification result with level slicing by 80m was better than the others. In this case, the size of co-occurrence matrix was 20*20 pixels.

  4. A combination of the homogently and correlation gave better results that the others for classification of drainage systems.

Fig. 5 Effects of window size


Fig. 6 Effects of Grid Size


Fig. 7 Effects of Quantization Level

Comparison with Drainage Systems
A plain, dendrite drainage, parallel drainage in gentle area and parallel drainage in steep area were selected as the study area in SO area. And parallel drainage in steep area was selected as those in those in KITAMI area.

Fig. 8 and Fig.9 show the histograms of homogeneity and correlation in those areas respectively. As shown in fig 10, drainage systems were classified by the parallel piped classification method.


Fig. 8 Histogram of Homogeneity


Fig. 9 Histogram of Correlation

Fig. 10 Threshold of Classification

Fig. 11 and Fig. 12 shows the results of the classification by experts and that by texture analysis respectively. The mapping discrepancies between computer generated pattern and expert's made pattern were shown in Table 2.


Fig. 11 The Classification by Experts


Fig. 12 The Classification by Texture Analysis

Table 2 Discrepancies of Classification in KITAMI Area
drainage systems Discrepancies
parallel drainage
dendritic drainage in gentle area
dendritic drainage in steep area
unknown		 47.8%
34.9%
34.9%
25.4%

Conclusions
  1. Drainage pattern classification has been developed by the authors using texture analysis. A combination of correlation and homogeneity gave the best results as compared with other types of texture. Computer-generated drainage patterns would be very useful for geological survey.

  2. There are some differences between the classified image by texture analysis and that by experts. Such differences exist mainly in the areas where DEM is not correct. If DEM is correct, it would be expected that classification accuracy of drainage systems would become higher.

Reference
  • Japan Association of Remote Sensing. "Processing and Analysis of Images", 1986