At present,most deep learning-based building extraction is based on semantic segmentation without considering building geometric characteristics,while traditional methods only consider their grayscale features in remote sensing building extraction,both of which are difficult to extract effectively.To address this problem,spectral information is studied and a building extraction method that fuses spectral features and super pixels is proposed.The method firstly generates many subregions with different shapes and sizes of superpixels based on the watershed transform;then the spectral features of buildings are used to merge the superpixels of buildings to achieve the primary extraction of buildings from remote sensing images;on this basis,the extracted buildings are rejected,and then the sliding field operation is selected to suppress the noise according to the geometric features;finally,the buildings are post- extracted from remote sensing images according to the maximum inter-class variance(Otsu).Finally,the buildings are post-extracted according to the maximum inter-class variance(Otsu).The buildings of remote sensing images processed in this paper include four types of areas:wasteland,mountainous area,suburban area and urban area.By comparing and verifying with the classical algorithm,the experimental results show that the algorithm is superior in remote sensing buildings extraction.