An explainable Deep Q-learning and convolutional neural network framework for rice leaf disease detection

Abstract

Rice leaf diseases pose a significant threat to global food security by reducing crop productivity and causing substantial economic losses. The traditional diagnosis method is manual method, which is low in efficiency, subjective and not suitable for large-scale agricultural monitoring. Despite the advances in automated disease detection using deep learning methods like CNNs, GANs, and transfer learning models, these techniques remain highly computational, not very flexible, and struggle to perform well in different imaging scenarios. Considering these drawbacks, this paper introduces an Explainable Deep Q-Learning based CNN framework which employs CNN-based feature extraction and Deep Q-learning-based adaptive policy optimization for rice leaf disease classification. The proposed model continuously refines the classification actions through reward-based learning, which makes the model more robust in various agricultural imaging environments, in contrast to traditional supervised CNN models that have static classification decisions. The proposed model achieved 98.5% accuracy, 98.52% precision, 98.50% recall, and a 98.51% F1-score, outperforming existing CNN, GAN, reinforcement learning, and transformer-based methods. It also offers a high computational efficiency of 14.2 GFLOPs, 248 MB memory consumption, ~ 48 min of training time, and 6.8 ms inference time per image suitable for resource constrained applications in agriculture. The results demonstrate the effectiveness, scalability, and practical applicability of the proposed framework. The proposed framework performs well on benchmark datasets but more research in the deployment of the edge-devices under different real-world agricultural settings will be investigated in future work.

Code and data availability

The paper trains its Deep Q-CNN rice leaf disease classifier on public Kaggle rice leaf image datasets, which are cited with explicit public URLs and qualify as paper-specific phenotyping image inputs. The authors' own derived data/analysis artifacts are only available upon request, so no authors' code or trained model