Effective Building Extraction From High-Resolution Remote Sensing Images With Multitask Driven Deep Neural Network

Building extraction from high-resolution remote sensing images has widely been studied for its great significance in obtaining geographic information. Many methods based on deep learning have been tried for the task; however, there is still much to explore about designing layers or modules for remote sensing data and taking full use of the unique features of buildings like shape and boundary. In this letter, an end-to-end network architecture based on U-Net is proposed. The U-Net architecture is modified with Xception module for remote sensing images to extract effective features. Also, multitask learning is adopted to incorporate the structure information of buildings. Two standard data sets (Massachusetts building data set and Vaihingen Data set) of high-resolution remote sensing images are selected to test our model and it achieves state-of-the-art results.