Classification of Very High-Resolution Remote Sensing Imagery Using a Fully Convolutional Network With Global and Local Context Information Enhancements

摘要

Deep learning methods for semantic image segmentation can effectively extract geographical features from very high-resolution (VHR) remote sensing images. However, these methods experience over-segmentation in low-level features and a breakdown in the integrity of objects with fixed patch sizes due to the multi-scaled geographical features. In this study, a dual attention mechanism is introduced and embedded into densely connected convolutional networks (DenseNets) to form a dense-global-entropy network (DGEN) for the semantic segmentation of VHR remote sensing images. In the DGEN architecture, a global attention enhancement module is developed for context acquisition, and a local attention fusion module is designed for detail selection. This network presents the improved semantic segmentation performance of test ISPRS 2D datasets. The experimental results indicate an improvement in the overall accuracy (OA), F1, kappa coefficient and mean intersection over union (MIoU). Compared with the DeeplabV3 and SegNet models, the OA improves by 2.79 and 1.19; the mean F1 improves by 3.43 and 0.88; the kappa coefficient improves by 4.04 and 1.82; and the MIoU improves by 5.22 and 1.47, respectively. The experiments showed that the dual attention mechanism presented in this study can improve segmentation and maintain object integrity during the encoding-decoding process.