MMTA-ACDD: Multi-teacher online adaptive learning and multimodal augmentation synergy for robust crop disease detection

The diagnosis and control of crop diseases are crucial for ensuring agricultural productivity. Single-stage real-time efficient detection models have been widely applied to crop disease detection tasks, but typically perform well only when the training and test sets have consistent distributions. However, in practical crop cultivation environments, real-time detection models not only need to perform excellently on source domain data but also must adapt to target domain data with different distributions. To address this challenge, this paper proposes a primary–auxiliary teacher collaborative enhanced supervision test-time adaptation framework to improve the robustness of real-time detection models in dynamic crop cultivation environments. Specifically, first, we propose a primary teacher supervision and auxiliary teacher guidance strategy, which combines weak augmentation and multimodal background strong augmentation techniques, respectively, to generate higher-quality pseudo-labels for the student model, thereby alleviating the problem of error accumulation in adaptive training. Second, we propose a maximum gradient dynamic adaptive recovery strategy(MGDA), which retains more effective knowledge from the source model by setting a maximum recovery trigger threshold and dynamic adaptive mechanism, preventing the student model from forgetting original knowledge when adapting to new domains. Finally, to enhance the robustness of the multi-teacher architecture under target distribution shift, we introduce a target class-aware contrastive learning strategy(OACL) to more effectively utilize pseudo-labels for feature learning in crop disease target detection tasks. Experimental results demonstrate that our proposed framework significantly improves the performance of real-time detection models in dynamic environments, providing an effective solution for crop disease detection.