DeMamba: Deformable Local Feature Learning with State Space Semantic Reasoning for Crop Protection and Precision Agriculture

Agricultural scene understanding through semantic segmentation is fundamental for precision agriculture, enabling phenological analysis, yield estimation, disease assessment, and crop management. Despite recent advances in CNN-, transformer-, and Mamba-based architectures, existing methods still exhibits inherent limitations in balancing fine-grained local details and global semantic reasoning. We propose DeMamba, a novel architecture that integrates deformable convolution-based local feature modeling with efficient Mamba-based global context modeling. Specifically, we introduce a Global Semantic State Validation (GSSV) module to enhance long-range dependencies and an Adaptive Local–Global Fusion (ALGF) gated mechanism for dynamic feature integration. Furthermore, a two-stage training strategy with gate supervision and structure-aware stage loss is employed to improve segmentation consistency. Extensive experiments demonstrate that DeMamba consistently outperforms existing SOTA methods on a challenging UAV-based rice panicle dataset (+1.00 IoU, +2.13 Acc, and +5.04 to +3.36 IoU across tiny-to-large panicles), while also improving performance on additional agricultural benchmarks, confirming its robustness and generalization.