LViM: Language-Infused Visual Mamba for apple leaf pests and diseases precise segmentation in complex environments

Apple leaf disease segmentation is critical for yield and quality preservation in what is globally one of the most economically significant fruit crops. Despite recent advances in deep learning, real-world orchard environments present three primary challenges: (1) low contrast between lesions and background textures, which hinders accurate localization; (2) leaf overlap and occlusion, leading to incomplete feature representation and increased false negatives; and (3) the inherent limitations of unimodal RGB imagery in capturing subtle pathological features, which constrains generalization and accuracy. To address these issues, we proposed Language-Infused Visual Mamba (LViM), a dual-path U-Net architecture that integrates Mamba and Transformer modules for semantic-visual feature fusion. LViM achieves robust segmentation in complex environments through three core innovations: (1) A U-shaped Multimodal Transformer (MTT) branch integrated with AMBERT, which leverages inter-modal semantic relationships to enhance textual feature extraction and provide high-level semantic cues, thereby improving lesion-background discriminability; (2) a U-shaped Visual State Space (VMamba) branch that employs 2D Selective Scanning (SS2D) and Visual State Space (VSS) blocks to capture global context and fine-grained details, mitigating the impact of occlusion; and (3) Cross-Attention Gate Fusion (CAGF) and Linguistic Cross-Nested (LCN) modules that facilitate efficient cross-modal alignment and hierarchical feature modeling to better identify subtle lesions. Experimental results demonstrate that LViM consistently outperforms the VM-UNet baseline, yielding improvements of 4.05% in Precision, 4.25% in Dice coefficient, 4.49% in mIoU, and 4.23% in Recall.