SUNet: Coffee Leaf Disease Detection Using Hybrid Deep Learning Model

Abstract

Leaf mining, rust, bacterial blight, and berry pathology are major diseases in coffee plants. These diseases not only reduce yield but also affect quality. Early detection and targeted treatment are crucial to mitigate their effects. This paper introduces an efficient hybrid deep learning model, SUNet, for prediction and classification of healthy and diseased coffee leaves. SUNet integrates U-Net with Segnet’s encoding system, using VGG16 for robust feature extraction. A decoder with skip connections is used to preserve spatial details. Mask R-CNN is also employed for instance segmentation, accurately localizing disease spots. A pyramid pooling module captures multi-scale contextual information. The model is tested using two benchmark datasets, JMuBEN and JMuBEN2. These datasets contain a wide range of coffee leaves affected by phoma, cercospora, or rust, along with healthy samples. SUNet achieved significant performance improvement over other models in terms of accuracy, Intersection over Union (IoU), F1-score, precision, and recall by 1.22%, 1.21%, 1.17%, 1.19%, and 1.24%, respectively. These improvements demonstrate that SUNet can be used for the early detection and classification of coffee leaf diseases. Therefore, with precise and timely interventions, SUNet can help farmers minimize crop losses, enhance coffee production quality, and reduce reliance on harmful chemical treatments.

Keywords

• Agricultural applications
• coffee leaf disease
• deep learning
• semantic segmentation
• SUNet
• SegNet