UAV image segmentation in complex mining areas by integrating spatially adaptive feature modulation and multi-scale context fusion modules

Abstract

To address the misclassification of visually similar ground objects and inaccurate boundary localization caused by background interference, as well as small-object omission, fragmented regions, and incomplete boundaries resulting from scale variations, this study proposes MASwin-Unet for semantic segmentation of complex mining areas. The model incorporates a Spatially Adaptive Feature Modulation Module (SAFM) to suppress background noise and similar-texture interference while enhancing key regions and boundary details. It also introduces a Multi-Scale Context Fusion Module (MSCF) to integrate contextual information across different receptive fields and improve multi-scale feature modeling. An experimental dataset was constructed from UAV imagery of the Dongshan Xiuqing Mining Area, and experiments were also conducted on AeroScapes and LoveDA. MASwin-Unet achieved an OA of 97.18%, a Kappa coefficient of 95.79%, an mIoU of 85.14%, and an mF1 of 90.73%, outperforming BEDSN, FADC, DeepLabV3 + , PSPNet, HRNet, SegFormer, TransUNet, U-Net, and Swin-Unet. The results demonstrate good generalization on public datasets and show that the model effectively reduces similar-object misclassification, small-object omission, and incomplete boundary segmentation in complex UAV scenes.