Generalization-Aware Deep Vision Framework for Rice Leaf Disease Recognition across Seasonal and Environmental Variability

The rice leaf diseases are very dangerous to the crop production especially under seasonal changes and environmental changes where the disease symptoms hold a high visual variation-diversity. Majority of the available deep learning-based disease recognition systems are trained in controlled or semi-controlled conditions and tend to miss generalization when the illumination, complexity of background, climate and crop growth stage vary. In order to overcome this weakness, this paper is proposing a generalization-conscious deep vision architecture in robust recognition of rice leaf disease at seasonal and environmental variations. The model presented is based on the learning of strong and invariant visual representations through the combination of adaptive feature learning, global contextual modeling and regularization-based training approaches. The framework enhances the resistance to changes in the domains that can be experienced in real-field settings by collectively capturing characteristics of local lesion and stable global feature patterns. The validity of the proposed approach in the context of multi-season rice leaf image datasets was proven through extensive experiments that showed that the approach was more effective in overall classification accuracy and robustness in cross-seasonal and cross-environmental evaluation conditions than the traditional deep learning models. These findings affirm the appropriateness of the framework to use in precision agriculture systems to ensure reliable deployment at scale, and in a practical manner.