VMamba for plant leaf disease identification: design and experiment

Introduction: The rapid spread of crop diseases poses a severe threat to agricultural production, significantly reducing both the yield and quality of crops. In recent years, plant disease recognition technologies based on machine vision and artificial intelligence have made significant progress. However, current mainstream deep learning architectures still face numerous challenges in detecting agricultural plant diseases. These include issues such as the complexity of agricultural environments and the reduced accuracy and increased training time caused by small sample sizes of agricultural plant diseases. Methods: To address these challenges, we introduce the VMamba visual backbone model into the task of detecting agricultural plant diseases. This model effectively reduces computational complexity through a selective scanning mechanism while significantly improving classification accuracy by maintaining a global receptive field and leveraging dynamic weighting advantages. Our study proposes the DDHTLVMamba method, which combines VMamba with diffusion models and transfer learning techniques, and applies it to the detection of plant diseases in small-sample agricultural datasets. This research evaluates the performance of VMamba across different datasets and training methods, conducting comparative analyses with mainstream deep learning architectures. Results and discussion: Experimental results demonstrate that the VMamba model outperforms mainstream models such as ResNet50, Vision Transformer, and Swin Transformer in disease recognition accuracy, whether on large-scale datasets like PlantVillage or optimized small-sample disease datasets, showcasing superior performance. Compared to Swin Transformer, VMamba achieves a 3.49% increase in accuracy while reducing training time by 80%. Furthermore, the proposed DDHTLVMamba training method demonstrates its effectiveness on small-sample datasets, significantly reducing pre-training time while maintaining recognition accuracy comparable to that achieved with large-sample transfer learning. This study provides an innovative approach for the efficient identification of agricultural diseases and is expected to advance the development of intelligent agricultural disease prevention and control technologies.