Performance Evaluation of Semantic Segmentation Architectures for Rice Leaf Disease Segmentation

Abstract

National food security is strongly influenced by rice productivity, which is vulnerable to various rice leaf diseases. Manual detection based on visual observation is often subjective and inefficient. This research proposes a deep learning-based semantic segmentation approach for automatically detecting rice leaf disease regions. Three semantic segmentation architectures, namely U-Net, PSPNet, and DeepLabV3+, were compared using the same ResNet34 backbone and identical training configurations to ensure objective evaluation. The dataset used was the Rice Leaf Disease with Segmentation Labels dataset, consisting of rice leaf images and pixel-level annotations for several rice leaf diseases. The research stages included data pre-processing, image augmentation, model training, and evaluation using Intersection over Union (IoU) and pixel accuracy metrics. Experimental results showed that DeepLabV3+ achieved the best performance with an IoU score of 0.9804 and pixel accuracy of 0.9235, outperforming U-Net and PSPNet. The findings indicate that semantic segmentation can provide more detailed and accurate detection of rice leaf disease regions and has the potential to support the development of artificial intelligence-based precision agriculture systems.