Decision tree algorithm of automatically extracting farmland information from SPOT-5 images based on characteristic bands

Farmland protection is crucial for China's sustainable development. Extraction of farmland information from remotely sensed images is significant for surveying, planning and protecting farmland, especially for basic farmland. In this paper, taking Tongxiang county as a case study area, SPOT-5 images collected on Auguest, 2002 were used. The approach to extraction was discussed. First, the spectral characteristics of farmland and other five land-use types in this area were analyzed to find the possibility of extracting farmland from the background. The results show it is difficult to distinguish farmland information from background on the SPOT-5 images because of complexity of spectrum and lack of band information. Second, taking those into account, characteristic bands for farmland extraction were proposed and merged into SPOT-5 images in order to increase spectral information and, at the same time, improve the separability. Third, a simple model of decision tree was applied to extract farmland information. Finally, the results were checked by visual and statistical accuracy assessment. The results suggest that the model based on characteristic bands is simple and effective, and the accuracy by the model is much higher than that by the supervised classification method. However, some pixels in the neighborhood area between farmland and mulberry were misjudged. The landscape type(other farmland) was mostly isolated and influenced by human activities to a high extent.