Built-up area mapping in China from GF-3 SAR imagery based on the framework of deep learning

Abstract

Timely and reliable acquisition of built-up areas (BAs) information in large regions is of great significance for the assessment of the influence of human activities in local, regional and even global environments and promotes the sustainable development of cities. In this study, a deep learning based framework is presented to automatically extract BAs from Gaofen-3 (GF-3) data in China. In the framework, a U-Net structure based deep learning with convolutional neural network is implemented for semantic segmentation to extract BAs. To overcome the problem of network degradation when the neural network layers become deeper, residual blocks were introduced via identity shortcut connections to add to U-Net. About 30,000 samples containing different distribution types of BAs and corresponding labeled binary images, which are from 27 scenes of GF-3 SAR images covering different regions (for example plain, hilly and mountainous regions) in China, were obtained for model training. With the trained model, more than 1700 scenes of 10-m resolution GF-3 Fine Stripmap II (FSII) mode single-polarized images covering the whole country of China were processed to obtain the BA map of China. On the one hand, quality evaluation based on the confusion matrix was performed to determine the accuracy of the results with the ground truth data, which were manually interpreted by experts. The experimental results show that among 34 provinces or regions, the overall accuracy ranges from 76.05% to 93.45%, and the F1-score ranges from 0.69 and 0.89. On the other hand, the results were also compared with other semantic segmentation methods, such as FCN, PSPNet and U-Net, and other global urban area products, such as the Global Urban Footprint (GUF), Finer Resolution Observation and Monitoring of Global Land Cover (FROM-GLC) and Global Human Settlement Layer (GHSL). The results show the effectiveness of the proposed method.