Multiscale Semantic Segmentation of Remote Sensing Images Based on Edge Optimization

Semantic segmentation of remote sensing images is crucial for disaster monitoring, urban planning, and land use. Due to scene complexity and multiscale features of targets, semantic segmentation of remote sensing images has become a challenging task. Deep convolutional neural networks capture remote contextual dependencies that are limited. Meanwhile, restoring the image size quickly leads to undersampling at object edges, resulting in poor boundary prediction. Therefore, this article proposes a multiscale semantic segmentation network of remote sensing images based on edge optimization, namely, multiscale edge optimization network (MSEONet). The decoder of the network consists of a multiscale context aggregation (MSCA) module, a coarse edge extraction (CEE) module, and an edge point feature optimization (EPFO) module. The MSCA module is used to capture multiscale contextual information and global dependencies between pixels. The CEE module is used for boundary extraction of multiclass coarse segmentation results. The EPFO module is used to optimize edge point features during the upsampling process. We conducted extensive experiments on the International Society for Photogrammetry and Remote Sensing (ISPRS) Potsdam 2-D dataset, the ISPRS Vaihingen 2-D dataset, and the FLAIR #1 dataset. The results show the effectiveness and superiority of our proposed MSEONet model compared to most of the state-of-the-art models. The CEE and EPFO modules can enhance the edge segmentation effect without increasing the computational and parametric quantities too much. The code is publicly available at https://github.com/HuangWBill/MSEONet.