Abstract:

It is difficult to acquire optical remote sensing images with good quality in Wolong nature reserve because of the weather conditions. Meanwhile, the terrain is very complex and the species are very rich, which leads to poor classification results when using Landsat TM data to study the panda's habitat. In this work, an effective classification method is proposed to improve classification accuracy. First, three features including greenness, brightness, and wetness are extracted from Landsat TM image by tasseled cap transformation. Second, the abundance index is deduced by matched filtering method. Then, the rules for decision tree are established by combining the results of the tasseled cap and matched filtering methods, and used to map the land use classification distribution in the study area. Finally, the classification results are validated by field measurements. The results show that the greenness is an effective measurement for extracting forest, wetness can distinguish between meadow and shrub, and brightness can improve the classification accuracy for snow. A matched filtering method for spectral unmixing removes image noise and effectively detects the spectra of targets. The overall accuracy of the decision tree classifier reaches 83.33%, which is 7.67% higher than that of the maximum likelihood classifier. This method improves the effectiveness and accuracy of image classification in Wolong nature reserve for giant pandas.

Key words: image clssification, tasseled cap, matched filtering, decision tree, giant panda habitat