Multiscale CNN-state space model with feature fusion for crop disease detection from UAV imagery

Accurate detection of crop diseases from unmanned aerial vehicle (UAV) imagery is critical for precision agriculture. This task remains challenging due to the complex backgrounds, variable scales of lesions, and the need to model both fine-grained spot details and long-range spatial dependencies within large field scenes. To address these issues, this paper proposes a novel Multiscale CNNState Space Model with Feature Fusion (MSCNN-VSS). The model is specifically designed to hierarchically extract and integrate multi-level features for UAVbased analysis: a dilated multi-scale Inception module is introduced to capture diverse local lesion patterns across different scales without sacrificing spatial detail; a Visual State Space (VSS) block serves as the core component to efficiently model global contextual relationships across the canopy with linear computational complexity, effectively overcoming the limitations of Transformers on high-resolution UAV images; and a hybrid attention module is subsequently applied to refine the fused features and accentuate subtle diseased regions. Extensive experiments on a UAV-based crop disease dataset demonstrate that MSCNN-VSS achieves state-of-the-art performance, with a Pixel Accuracy (PA) of 0.9421 and a mean Intersection over Union (mIoU) of 0.9152, significantly outperforming existing CNN and Transformer-based benchmarks. This work provides a balanced and effective solution for automated crop disease detection in practical agricultural scenarios.