Multi-scale mamba attention and residual learning for robust agricultural image segmentation in precision agriculture

Accurate segmentation of agricultural images is critical for precision agriculture, enabling efficient crop management, weed detection, and resource optimization. We propose a novel segmentation network tailored for agricultural imagery, integrating multi-scale mamba attention (MSMA) Blocks and Conv2D residual blocks within a U-Net-like architecture to address the challenges of variable scales, complex boundaries, and heterogeneous object shapes. The MSMA Blocks enhance multi-scale feature extraction, while Conv2D residual blocks improve feature propagation and spatial consistency. Tested on the Agriculture-Vision dataset, our model outperforms state-of-the-art methods like AgriSegNet and DeepLabV3 + in terms of Mean Intersection over Union (mIoU), particularly in challenging classes such as Planter Skip and Weed Cluster. An ablation study further demonstrates the critical contributions of both MSMA Blocks and residual connections to the network’s performance, making our approach a promising solution for precision agriculture tasks.