<![CDATA[Rice leaf diseases represent a critical threat to paddy productivity and national food security in Indonesia, while field-level diagnosis still relies heavily on subjective visual inspection. This study aims to improve the classification performance across five rice leaf image classes (Bacterial Leaf Blight, Tungro, Normal, Brown Spot, and Blast) through the modification of the ResNet50 architecture and data augmentation techniques. A public Kaggle dataset comprising 6,034 images was utilized, incorporating four training augmentations: Random Zoom, Random Brightness, Horizontal Flip, and Vertical Flip. Eight experimental scenarios (S1–S8) were systematically formulated to evaluate ResNet50 with a Multilayer Perceptron (MLP) classification head against ResNet50 without MLP, tested under baseline, single augmentation, and combined augmentation settings. Performance was evaluated on 250 stratified test images using confusion matrices with accuracy, precision, recall, and macro-averaged F1-score metrics. The results demonstrate that modifying the architecture by eliminating the MLP classification head simplified the model complexity and increased the baseline accuracy from 88.40% (S1) to 89.20% (S3). Data augmentation consistently enhanced performance across all scenarios, with Random Zoom emerging as the most impactful individual technique (92.40% accuracy). The optimal configuration was achieved in scenario S8 (ResNet50 without MLP combined with all four augmentation techniques), attaining an accuracy of 94.40%, precision of 94.68%, recall of 93.84%, and F1-score of 94.23%, representing a 6.00 percentage point improvement over the baseline model. These findings confirm the efficacy of combining architecture streamlining and diverse geometric-photometric transformations for robust rice leaf disease diagnosis.]]>
