Técnicas de detección de enfermedades del tomate en condiciones climáticas no estacionarias

Tomato growth is highly susceptible to diseases, making accurate identification crucial for timely intervention. While deep learning models like the YOLO family have demonstrated success in detecting diseases in agricultural settings, they typically assume that training and testing data are independently and identically distributed (i.i.d.), which often doesn't hold in real-world scenarios. When pre-trained models are applied to new environments, performance can degrade due to domain shifts. To address this, we propose CTTA-DisDet, a continuous test-time domain adaptation framework for tomato disease detection that adapts models to evolving environments during testing, improving generalization in unseen domains. CTTA-DisDet utilizes a teacher-student architecture where both models share the same structure. Dynamic data augmentation is introduced, involving explicit and implicit augmentations. Explicit augmentation corrupts input images, while implicit augmentation uses large language models (LLMs) to generate new domain data. The teacher model learns generalized knowledge, and the student model mimics the teacher to distill domain-specific information. During testing, pseudo-labels generated by the teacher update the student model. To prevent catastrophic forgetting, a subset of neurons is randomly restored to their original weights during each test-time iteration. The teacher model is continuously updated via exponential moving average (EMA). Experimental results demonstrate that CTTA-DisDet achieves an impressive 67.9% performance in continuously changing cross-domain environments, significantly benefiting practical applications in non-stationary settings.