MSCSCNet: Multiscale Convolution Selective Context Network With Boundary Supervision for Semantic Segmentation of Remote Sensing Images

Semantic segmentation of remote sensing images remains a challenging task due to the presence of small and indistinct objects, as well as semantic ambiguities arising from occlusions and similar visual appearances. Furthermore, the downsampling operations employed in backbone networks often lead to the loss of crucial boundary details, particularly in densely packed objects. To address these challenges, we propose a multiscale convolution selective context network (MSCSCNet) with boundary supervision. The MSCSCNet is designed with three key modules to enhance segmentation performance. First, the automatic selective context module adaptively selects appropriate receptive fields for objects of varying scales through a context selection mechanism, thereby minimizing semantic inconsistencies. Subsequently, the spatial semantic relationship module enables multiscale feature fusion by capturing the correlation between semantic and spatial information. Finally, the boundary feature extraction module reproduces highly precise boundary details, resulting in refined segmentation outcomes. Extensive experiments underscore the effectiveness of MSCSCNet across three benchmark datasets, with mean intersection over union scores of 70.22% on the iSAID aerial image dataset, 53.99% on the LoveDA domain adaptation dataset, and 80.24% on the ISPRS Potsdam high-resolution remote sensing dataset.