Discussion
Table 6 shows that all three methods described earlier can extract urban built-up lands from the new three-thematic-band images with high overall accuracy (on an average close to 95 percent in the 12 tests). The best overall accuracy, 98.5 percent, was achieved through the logic calculation method, which is the fastest, easiest, and most objective one among the three methods. One simple logic statement can generally achieve a quite good extraction result like in Quanzhou case. Nevertheless, one more condition added can
considerably increase extraction accuracy such as in the Fuzhou case, as it well discriminate vegetation from built-up land. Consequently, no confusion was found between them in the sampling procedure of accuracy assessment (Table 6). The threshold value used in the conditional statement was just simply taken from the maximum of the built-up land class in SAVI-band and did not need to bother to find it. The method
also has the highest average overall accuracy (94.98 percent) among the three considered. The PC2 can also get high accuracy up to 97 percent. This is probably owing to the fact that the contrast between
built-up land and vegetation or water in the PC2 image is further enlarged as indicated by their contributions to PC2 and by opposite signs of the loadings (Table 5). Therefore, the method can effectively enhance built-up land features while depressing water and vegetation information in the PC2 image. The conventional supervised classification on the new three-band image can produce very high accuracy up to 98 percent, while the same classification method performed on the original image can only have maximum accuracy of 81.5 percent. The good separation of the spectral clusters of the three urban categories in the new three-band image (Figure 4) greatly reduced the confusion between them, and thus considerably improves the classification accuracy. The supervised classification on the raw and DOS-corrected original seven-band images got the identical results because the DOS method is only shift the origin of the dataset as indicated by Song et al. (2001). The accuracy based only on the NDBI image using a default threshold value of 0 is lower than 85 percent, the minimum acceptable overall accuracy proposed by Anderson et al. (1976). Step-by-step adjusting the threshold value can reach higher overall accuracy but never exceeds 88.5 percent.
The confusion matrix in Table 6 shows that the error is obviously caused by high overestimation of the non-built-up lands. Visual inspection of extraction result (Figure 6e and 6i) can find that the confusion mainly lies between built-up land and vegetation classes. As discussed earlier, besides built-up lands many vegetation areas also have positive values in NDBI imagery, and thus made the confusion with the built-up land areas. This directly resulted in the low accuracy of the extraction. It seems no significant different between the extraction results from raw and DOS-corrected imagery and no rule can be found within them. In Quanzhou’s case, the average accuracy of raw image is slightly greater than that of DOScorrected image (94.0 percent versus 93.3 percent). While the situation is changed in Fuzhou case, and the average accuracy of DOS-corrected image is 0.5 percent higher than that of raw image (96.17 percent versus 95.67 percent)