<![CDATA[Chili leaf disease reduces yield and farmer income in Indonesia, yet automated diagnosis is typically evaluated using a single architecture or artificially balanced datasets that do not reflect field conditions. This study compared eight CNN architectures (ResNet50, ResNet101, EfficientNetB0, EfficientNetB3, MobileNetV2, MobileNetV3Large, VGG16, and DenseNet121) for six-class chili leaf disease classification on an original dataset of 1,847 images, using on-the-fly augmentation only on the training subset. Each architecture was trained with a frozen feature-extraction transfer-learning protocol and class weighting across three random seeds (24 runs total) to assess robustness. MobileNetV3Large achieved the highest mean accuracy (98.80%) and was selected as the final model, reaching 98.92% accuracy, a macro-average F1-score of 98.09%, and a weighted-average F1-score of 98.90%. A Friedman test confirmed that performance differences among the eight architectures were statistically significant ( = 19.432, = 0.0069), with a Kendall's W of 0.9253 indicating strong ranking agreement across seeds. Grad-CAM visualization showed that the final model's predictions were consistently driven by symptomatic leaf regions rather than background artifacts. These results indicate that a lightweight, lower-parameter architecture can match or exceed deeper models under frozen feature extraction, offering a favorable performance-to-size trade-off for agricultural extension services. This study contributes a statistically validated benchmarking protocol enabling fair comparison across CNN design philosophies on a single, non-synthetically-balanced dataset, a setting most prior chili leaf disease studies have not provided.]]>
