There are various objects on the earth, for which traditional classification methods can hardly considering their diversity and specific characteristics; therefore, it is difficult to obtain stable and precise classification results. Under this circumstance, an adaptive extraction method based on spectral matching, is proposed to extract thematic object completely and accurately from remote sensing image. This algorithm first selects the endmember of thematic object, and use spectral matching method to get the matching index; then, histogram threshold based auto-segmentation method is used to initially separate thematic object from background; next, thematic object pixels are searched out and taken as seed points to proceed region growing to gain the local area of thematic object; last, the boundary of the local area is searched and iterative classification within the local area is employed to precisely extract thematic object’s precise partition. Experiments on ETM image to extract water and bareland are employed here, and through comparison with maximum likelihood classification and support vector machine (SVM) classification, it verifies the effectiveness and commonality of this method.