Multi-branch convolutional neural network for built-up area extraction from remote sensing image

Built-up area is one of the most important objects of remote sensing images analysis, therefore extracting built-up area from remote sensing image automatically has attracted wide attention. It is common to treat built-up area extraction as image segmentation task. However, it's hard to devise a handcrafted feature to describe built-up area since it contains many non-built-up elements, such as trees, grasslands, and small ponds. Besides, built-up area corresponds to large size local region without precise boundary in remote sensing image so that the precision of segmentation in pixel level is not reliable. To cope with the problem of built-up area extraction, a segmentation framework based on deep feature learning and graph model is proposed. The segmentation procedure comprises of three steps. Firstly, the image is divided into small patches whose deep features are extracted by the devised lightweight multi-branch convolutional neural network (LMB-CNN). Secondly, a patch-wise graph model is constructed according to the learnt features, and then is optimized to segment built-up area with patch-level precision in full frame of remote sensing image. At last, post-processing step is also adopted to make the segmentation result visually intact. The experiments verify that the proposed method shows excellent performance by achieving high overall accuracy over 98.6% on Gaofen-2 remote sensing image data with size of 10,240 x10,240. (C) 2019 Elsevier B.V. All rights reserved.