CFAEVM-UNet: cross-fusion attention enhanced VM-UNet for UAV image segmentation of small and dense field crop pests

Abstract

Precise segmentation of small and dense field crop pests in UAV images remains challenging due to limited long-range dependency modeling in CNNs and quadratic complexity in ViTs. Although VM-UNet captures global context efficiently, it lacks local texture detail and multi-scale feature integration for small targets. To address these limitations, this paper proposes CFAEVM-UNet (CrossFusion Attention Enhanced VM-UNet), which incorporates three core modules: a Visual State Space (VSS) block for efficient long-range dependency modeling, a Local-Global Spatial-Channel Gated Attention (LG-SCGA) module for multi-scale gated attention fusion, and a Multi-Stage Channel-Wise Attention (MSCWA) module for cross-stage channel reweighting. Extensive experiments on an integrated dataset (IP102 subset combined with UAV-captured field images) demonstrate that CFAEVM-UNet achieves state-of-the-art performance, attaining an mIoU of 81.23% and a DSC of 83.67%, outperforming U-Net by 8.89% in mIoU. This work provides an effective and practical solution for automated pest monitoring in precision agriculture.