Abstract:Degradation of grassland is a complex and gradual process. In order to make it more capable for monitoring the changes in undergoing-degradation and not-being-completely-deserted grassland with remote sensing data, one of available ways is to identify spectral characteristics in response to the differences in grassland coverage by machine discerning. However, it is challenging since the differences are subtle in a TM / ETM image or other images with similar spatial and spectral resolution. This paper discusses the methods for a learning machine to discern the subtle differences from TM / ETM spectral data. The key technologies can be summarized as: replacing the grass cover (D) with the so called "weighted grass cover (Dw)" to solve the difficulties in the fitting between Normalized Difference Vegetation Index(NDVI) from TM/ETM spectral data and field measured D, designing some mathematical descriptors which are independent of image types and acquisition environment and discerning the indicative differences in response to Dw in a multi-descriptor space so that the error rate in the discerning will be surely reduced. Dw is a weighted sum of the D, the grass height (H) and the proportion of the poisonous in all grass (Dup). With the use of Dw, The regression equations between Dw, calculated by D, H and Dup measured in the field, and the normalized vegetation indexes, such as NDVI, SAVI, etc., calculated with remote sensing spectral data will fit much better than that with the replacement of D for Dw. In order to improve the sensitivity to the differences in grass coverage by the learning machine, eight new mathematical descriptors, as the supplements of those vegetation indexes, have been proposed and then analyzed one by one in their correlation to Dw or tested in their validity for Dw segmentation. It is verified that some of the combinations of these descriptors are helpful for the learning machine to work out the laws of identifying pixels or cells belonging to different Dw. Otherwise, by comparing the classification accuracy in the use of two kinds of learning machines, i.e. DT and SVM, the types of learning machines are not so crucially to be able to decide discrimination accuracy as mostly desirable and, on the contrary, the construction of search space may be a more decisive factor. Our study shows that in the multi-descriptor spaces, the grasslands belonging to different Dw can reliably be classified with TM / ETM data. Being inspected by field sampling data, it is verified that the classification accuracy is nearer to or better than 80%. Based on the results in Dw classification, then the Dw changes can also be reliably deduced after adjusting the differences of luminance level between two images in different acquisition years.