MVMamba-Seg: a precise multimodal segmentation method for tomato disease monitoring via efficient Vision Mamba

Introduction Accurate segmentation of tomato leaf lesions is essential for disease diagnosis, severity assessment, and precision crop protection, yet field images contain complex backgrounds, diverse lesion appearances, irregular boundaries, and confusing foreground-background transitions. Methods We propose MVMamba-Seg, a text-guided multimodal Vision Mamba network that couples CNN and Vision Mamba streams with BERT-encoded expert descriptions for efficient global modeling. Dynamic Difference-Driven Feature Enhancement (DDFE) explicitly models foreground, background, and uncertainty branches, while the SSM Spatial Reweighting Attention Module (SSM-SRAM) refines skip connections and suppresses background noise. Results On a self-built dataset, MVMamba-Seg achieves 94.32% Dice and 89.26% mIoU using expert text; in cross-domain evaluation on PlantVillage, it achieves 81.82% Dice and 69.24% mIoU with 21.89M parameters. Compared with a Transformer-based counterpart, the Vision Mamba variant reduces parameters by approximately 31%, FLOPs by 42.8%, and improves inference speed by 60.1%. Discussion Through efficient global modeling with Mamba and the integration of expert disease descriptions with visual features, the framework delineates lesion contours and spatial distributions to support tomato disease monitoring and targeted crop protection.