Abstract: Mapping the information related to the number, area and location of rural buildings is the basis for rational planning of rural land and building beautiful and livable rural areas in the new era. However, due to the phenomenon that rural buildings are scattered and severely fragmented, it is still challenging to accurately extract the scattered rural buildings using remote sensing technology. In this study, we propose a rural building extraction method that combines object-oriented and deep confidence networks. Firstly, we use object-oriented scale segmentation based on the spectral, shape and texture features of rural buildings, and further use deep confidence networks to learn high-level semantic features such as texture and environment of different objects to extract rural building information. Compared with random forest commonly used image classification methods,the method in this paper performs better in rural building extraction, with clearer and more complete edge contours of the extracted patches, and better recognition of the distinction between the gap parts between different buildings, and less noise in the extraction results. The method can effectively and efficiently extract rural building information.

Key words: object-orientation, deep learning, rural buildings, high-resolution images