Vegetation and bare soil are the main land cover types in Karst and non-Karst area.Bare rock another land cover type only appearing in Karst area,is distributed extensively. There are a variety of methods to detect land cover in Karst area with remotely sensed imageries. But most ordinary remotely sensed approaches can't get the three attributes at the same time. In fact,they almost can do nothing in detecting the quantitative result of rock desertification. The linear spectral unmixing approach,based on the linear mixture model can estimate vegetation cover, bare soil and bare rock abundance at the sub-pixel scale. It's fit for studying land cover in Karst area because of the obtained quantitative unmixing results of vegetation cover, bare soil, bare rock and other types. In the case study of this paper, four endmembers were identified from two TM images of six bands taken on 17th February 1987 and 27th November 1999 respectively. They represent four land cover types: vegetation, bare soil, rock and shadow. The abundant distribution and RMS (Root Mean Square) distribution of the four land types were derived by linear spectral unmixing. The land cover changes could be learned from the multi-time comparison. During the period from 1987 to 1999, the proportion of bare soil reduced evidently due to the increase of vegetation cover and the improved cultivation management. While, the proportion of rocky desertification enhanced owing to the little amount of soil in some place. Bare soil reduced and rock increased mostly in rainfed cropland, grassland and economic woodland. Over-cultivating and overgrazing are still the main driving forces of rocky desertification in Karst area. So people still need to pay attention to water and water soil conservation in economic woodland because economic woodland mainly came from "dry land reforest". The linear spectral unmixing approach still needs to be improved, though it has been proved fitting for dealing with mixture pixel and detecting quantitative results of land cover by the case study and many other literature. If endmembers need to be identified from image, it's hard to detect those elements without extraordinary characters presenting in images of different bands. For example, urban region is just such an element difficult to detect. So the approach is not the most suitable method for urban region or rapid urbanized area. And the impact of shadow is still a problem, just as in other methods.