Rotation Equivariant Convolution-Based Network for Semantic Segmentation of High Spatial Resolution Remote Sensing Imagery

Deep Neural Networks (DNNs) have demonstrated strong feature extraction capabilities and achieved high accuracy in semantic segmentation tasks. However, they exhibit limitations in handling rotation variations, leading to inconsistent segmentation performances when processing remote sensing images containing variously oriented objects. To address this problem, a novel rotation equivariant convolution-based network (RECNet) is proposed for semantic segmentation of remote sensing images in this letter. In RECNet, a dual-branch encoder architecture is developed, where a CNN-based branch is designed to extract primary features, while a rotation equivariant convolutional branch is designed to capture robust rotation features. Experimental results on ISPRS Potsdam and Vaihingen datasets demonstrate that the proposed method can effectively handle rotation variations, improving segmentation accuracy and robustness than methods that without rotation equivariant convolution, with OA values of 89.03% on the Potsdam dataset and 90.76% on the Vaihingen dataset. The code of the proposed RECNet is available at https://github.com/LiYu616/RECNet.