LAUSR

As the central area of human activities, built-up area has been one of the most important objects that are recognized from a remote sensing image. Built-up area in different regions has characteristics as follows: the structure and texture of the built-up area are complex and diverse; the buildings have multitudinous materials; the vegetation distribution and background around the built-up area are changeable. The existing built-up area detection methods still face the challenge to achieve favorable precision and generalization ability. In this paper, a double-stream convolutional neural network (DSCNN) model is proposed to extract the built-up area automatically, which can combine the complementary cues of high-resolution panchromatic and multispectral image. Some postprocessing steps are adopted to make the results more reasonable. We manually annotated a large-scale dataset for training and testing DSCNN. Experiments demonstrate that the proposed method has a higher overall accuracy as well as better generalization ability compared to the state-of-the-art techniques.