Saurabh Bilgaiyan

Accurate identification of rice leaf diseases is essential to securing agricultural productivity and mitigating crop losses. Manual approaches are often inefficient and unreliable, particularly in large-scale farming. Although deep convolutional neural networks such as DenseNet have been applied to this task, their default configurations may not fully capture fine-grained disease features. This study aims to develop a series of enhanced DenseNet models that incorporate architectural improvements and optimized learning parameters to achieve highly reliable classification of rice leaf pathologies. We implemented baseline and modified versions of DenseNet121, DenseNet169, and DenseNet201, integrating Squeeze-and-Excitation (SE) blocks to enhance channel-wise feature calibration. The proposed approach is evaluated on a publicly available dataset comprising 3,829 rice leaf images distributed across six classes, including Brown Spot, Sheath Blight, Leaf Scald, Bacterial Leaf Blight, Leaf Blast, and Healthy rice leaves. To improve generalization and convergence, the modelswere fine-tuned usingKeras Tuner with a focus on optimizing the number of dense units, dropout rates, and learning rates. The proposed hybrid framework combines Squeeze-and-Excitation–enhanced DenseNet architectures withKerasTuner-based hyperparameter optimization, enabling joint feature refinement and systematic model optimization, which distinguishes it from existing DenseNet-based rice leaf disease detection approaches. The evaluation framework included dimensionality reduction techniques (PCA, t-SNE) and various statistical plots (histogram, KDE, box, and violin). Model performance was assessed using accuracy, precision, recall, F1-score, area under the ROC curve, and Cohen’s Kappa coefficient. All evaluated DenseNet-based models achieved consistently high performance, with accuracy, precision, recall, and F1-score values close to 0.99, while the Modified DenseNet-201 model yielded the highest overall results across all metrics. Its predictions exhibited strong confidence with minimal uncertainty, as evidenced by clear bimodal probability distributions and minimal misclassification in confusion matrices. The training history indicated smooth convergence with no significant overfitting. Notably, the Cohen’s Kappa score reached 0.9937, confirming excellent consistency beyond chance. The inclusion of SE blocks was especially effective in disambiguating diseases with similar visual traits. The proposed modifications to DenseNet architectures, supported by targeted hyperparameter tuning, significantly elevate performance in rice leaf disease classification. The models developed in this work demonstrate robust accuracy, strong interpretability, and practical viability for deployment in precision agriculture systems aimed at early disease detection.