ACMNet: An attention-convolution mixed neural network for semantic segmentation of high-resolution remote sensing images

Semantic segmentation of remote sensing imagery, which is widely used in land cover classification, urban planning, and environmental monitoring, has become a research hotspot in the fields of remote sensing and computer vision. As its applications continue to expand, improving the segmentation accuracy and generalization have become pressing challenges. However, most existing semantic segmentation models rely solely on either attention mechanisms or convolutional neural networks (CNNs), thus leaving room for performance improvement. In this study, we propose ACMNet (Attention-Convolution Mixed Network), which adopts a hybrid attention-convolution architecture. The decoder of ACMNet employs an innovative mixed design, where two branches are used to extract and integrate different types of contextual information and achieve hierarchical aggregation of contextual features, thereby enhancing its performance. To evaluate the performance of ACMNet, we conducted ablation experiments and comparative experiments on the Potsdam, Vaihingen, and WHU Building datasets. The results show that ACMNet achieved the best overall performance across all three datasets, with mF1 scores of 91.79%, 88.70%, and 96.56% on the Potsdam, Vaihingen, and WHU Building datasets, respectively. These findings indicate that ACMNet delivers superior performance in the semantic segmentation of high-resolution remote sensing images in urban areas, which provides valuable support for diverse tasks related to fine-grained urban management and dynamic monitoring.