<![CDATA[Rice leaf diseases pose a significant threat to agricultural productivity, and accurate automated detection is essential for timely intervention. This study presents a comparative evaluation of lightweight convolutional neural network architectures for the classification of six rice leaf disease categories: Bacterial Leaf Blight, Brown Spot, Leaf Blast, Leaf Scald, Narrow Brown Leaf Spot, and Rice Healthy. MobileNetV2 is proposed as the primary model and benchmarked against EfficientNetB0 and NASNetMobile. All three architectures were trained under an identical experimental setup comprising a two-stage transfer learning strategy, a unified custom classification head consisting of Global Average Pooling, Batch Normalization, two dense layers with dropout and L2 regularization, and a Softmax output layer. The dataset comprised 1,920 images across six classes obtained from Roboflow Universe, with no pre-augmentation applied by the original source. Training-time augmentation including rotation, shifting, shearing, zooming, and horizontal flipping was applied exclusively to the training subset. Experiments were conducted on a stratified split of 1,536 training, 192 validation, and 192 test images with a fixed random seed of 42 to ensure reproducibility. MobileNetV2 achieved the highest test accuracy of 96.35% and macro F1-score of 96.35%, outperforming EfficientNetB0 at 94.27% and NASNetMobile at 89.06%. In terms of computational efficiency, MobileNetV2 also demonstrated the most favorable deployment profile with a TensorFlow Lite model size of 2.75 MB and inference latency of 3.22 ms per image, indicating potential suitability for resource-constrained deployment scenarios. These results suggest that MobileNetV2 offers a competitive balance between classification accuracy and computational efficiency for rice leaf disease identification.]]>
