EGCM-UNet: Edge Guided Hybrid CNN-Mamba UNet for farmland remote sensing image semantic segmentation

Abstract

Segmenting farmland images is challenging due to their high color similarity to the background and irregular shapes, resulting in over/undersegmentation. To tackle these challenges, we propose the Edge Guided Hybrid CNN-Mamba UNet (EGCM-UNet) and design the oriented residual convolutional edge branch (ORCEB) to mine prior edge information. Additionally, the model designs a MaUNet module, which introduces the Visual State Space (VSS) block fused with Mamba to manage long-distance dependencies of image features, and uses the Edge-Guided Semantic Aggregation Module (EGSAM) for precise segmentation by fusing edge features with the VSS block’s output. Lastly, comparative experiments were conducted using selected baseline models on the AgriculturalField-Seg dataset. The results show that EGCM-UNet outperformed U-Net with a Mean Intersection over Union (mIoU) of 0.394 vs. 0.379 on the test set. This indicates the proposed model delivers good performance in the semantic segmentation task of farmland remote sensing images.