Classification of Durian Foliar Diseases by Xception and Mask R-CNN Models

Durian is an important export product in Thailand, with demand steadily increasing. However, durian farmers still face significant challenges in increasing productivity due to foliar diseases. This study applied deep learning methods to classify durian leaves into four categories: agal spot leaf, blight leaf, other diseased leaf, and healthy leaf. A dataset of 12,960 images was compiled from local farms and public sources, reflecting real-world field conditions. Two efficient image classification models, Xception and Mask R-CNN, were employed with carefully tuned hyperparameters. The ResNet-9 model was compared with the two proposed models. The results showed that both the Mask R-CNN and Xception models outperformed ResNet-9 on the same dataset. The Mask R-CNN achieved a high accuracy of 99.61%, precision of 97.80%, recall of 98.81%, F1 score of 98.30%, and a loss of 0.0987, when using the Adam optimizer with a maximum learning rate of 0.001 over 6,740 iterations. However, the training time of the Mask R-CNN was 377 seconds higher than the Xception model. The proposed models offer a capable tool for early disease detection and smart monitoring of durian leaf diseases, potentially supporting yield improvement and reducing dependency on manual diagnosis.