Mapping of Land Degradation from ASTER Data: A Comparison of Object-Based and Pixel-Based Methods

Abstract

Land degradation in Tongyu County, Northeast China was mapped from the visible, near infrared, and shortwave infrared bands of ASTER data using the per pixel-based maximum likelihood and object-based image classification methods, comparatively. In both methods the land covers were mapped into nine categories, three of which were related to land degradation. It is found that the ASTER image of 15 m spatial resolution allowed the mapping to be achieved at an overall accuracy of 70.6% using the pixel-based method. The accuracy for degraded land was slightly higher at 73.3%. If mapped from the same image segmented at 10 pixels (150 m) using the object-oriented method, the overall accuracy rose to 74.2%. However, the accuracy of severely degraded (i.e., bare ground) decreased, and the accuracy of degraded land also decreased to 65.8%. The overall accuracy rose to 76% if the classification was performed to the same image segmented at 20 pixels. However, the accuracy for degraded land was lowered further to 64.6%, even though the accuracy of bare ground was improved to 82.1%. It is concluded that object-oriented image classification does not fare much better than pixel-based image classification in mapping degraded lands from moderate-spatial-resolution satellite data such as ASTER due to their fragmented and discontinuous spatiality. At the 15 m resolution level, scale does not seem to exert a noticeable impact on the object-based classification accuracy.