<![CDATA[Rice (Oryza sativa) productivity is frequently threatened by foliar diseases such as Bacterial Leaf Blight, Brown Spot, Blast, and Tungro, which are often visually indistinguishable. This study achieved a high classification accuracy of 97.05% in detecting these diseases by optimizing the ResNet50 architecture with five optimizers Adam, Nadam, Adamax, RMSprop, and SGD and identifying Adamax as the most effective. Using transfer learning with ImageNet weights and data augmentation, the model was trained and validated on 4,400 labeled images from Kaggle, partitioned in a 70:20:10 ratio for training, validation, and testing. The methodological framework integrates three layers of innovation: (1) optimizing a deep residual CNN with comparative adaptive and non-adaptive optimizers; (2) employing transfer learning to accelerate convergence and reduce overfitting; and (3) deploying the best-performing model into an Android-based mobile application for real-time field detection. Results demonstrate that adaptive optimizers substantially enhance ResNet50’s learning stability and generalization compared to traditional methods. The Adamax variant exhibited the most stable convergence and minimal validation loss, proving effective for fine-grained visual differentiation between similar disease patterns. This research advances the current state-of-the-art in agricultural image classification by providing a systematic optimizer evaluation within a CNN transfer learning framework and extending its practical usability through mobile deployment. Future studies should address model compression, real-time inference optimization, and cross-crop generalization to strengthen the scalability of AI-assisted disease diagnosis in precision agriculture.]]>
