3. Data Processing

3.1 Processing of RADARSAT Data
The RADARSAT data were processed from Level-0 by Vexcel SAR Processor (VSARP) and single-look power images with 6.25 meters ground resolution were generated. Then the power images were filtered using mean filter with 7 by 7 moving window and finally converted into 8-bits image data in which power level is represented in dB. All RADARSAT and SPOT images were overlaid onto the topographic map with 1:25,000 scale. As RADARSAT images are much distorted by foreshortening due to topography, the digital elevation model (DEM) with 50 meters spatial resolution issued by Geographical Survey Institute (GSI) of Japan was used to correct foreshortening of RADARSAT images.

Figure 2 shows the temporal changing patterns of RADARSAT backscatter of some typical land cover types. As the RADARSAT backscatter is not calibrated perfectly, the figure shows it by relative power level in dB. In Figure 2 only rice shows a significant change of backscatter due to the change of its surface condition described in the previous section, and this suggests that rice-planted areas are possible to be extracted using backscatter change in multi-temporal RADARSAT images.

Fig.2 Temporal changes of RADARSAT-FIF backscatter of several land cover types.
3.2 Extraction of Rice-Planted Areas
Rice-planted areas were extracted using two methods from RADARSAT images. One was thresholding and the other was supervised classification by maximum likelihood (ML) classifier. The thresholding is a simple one that is, a pixel is extracted as rice if the backscatter change of that pixel is larger than a certain threshold value. The data combinations for thresholding were three, the first one was Apr.8 - May 26 -June 19 and the third the combination of the first and the second. For the third case, rice-planted areas are extracted by AND operation of the areas by two data combinations (Apr.8 May 26-June 19).

The second method, ML classifier, conducted land cover classification including rice category using the merged images of three or two temporal RADARSAT images, namely three or two channel images in which each channel indicates each temporal image. The three channel means to use all three temporal images and the two channel means to use only Apr.8 and May 26 because the backscatter change of rice is larger in the first two images than in the last two images in Fig. 2. For SPOT/HRV, the same method, ML classifier, was used to classify several land over types including rice.

4. Experimental Result

4.1 Result of Rice-Plated Area Extraction
Figure 3 shows the examples of rice-planted area extraction by RADARSAT, SPOT and finally by NLDI-landuse in the same area as Finger 1. It should be noted that NLDI-landuse indicates landuse as rice-fields and does not mean that they are always planted. In addition, the data were generated in an older time, therefore, some rice-fields are possible to change into other landuses. Top-left result was obtained by the thresholding of Apr.8 and May 26 with threshold value of -3dB in RADARSAT backscatter. The top-right was obtained by the ML classifier of all three temporal RADARSAT images. The classification score for rice by RADARSAT was about 92 percent for both of two and three channel, while that of SPOT was almost the same as RADARSAT.

RADARSAT Thresholding Apr-May :- 3dB



RADARSAT ML-classifier Apr.&May&June



SPOT/HRV ML-classifier June 21



NLDI-landuse <rice-field>
Fig. 3 Results of rice-planted area extraction by RADARSAT (top-left and top-right), by SPOT (bottom-left), and by NLDI-landuse (bottom-right)
In Figure 3, the rice extracted areas in both of the two results by RADARSAT seem quite similar to those in the result by SPOT,. while those by RADARSAT and SPOT are much smaller than those by NLDI-landuse. One of the reasons for the difference is considered to be caused by the time difference between the two images described already, and the other is considered to be due to rought spatial resoluation (100 meters) of NLDI-landuse.