A Deep Learning-Based Plant Disease Detection and Classification for Arabica Coffee Leaves

Coffee leaf disease is a growing concern to coffee agroforestry, predominantly caused by pathogenic fungi and, to a lesser extent, bacteria and viruses reducing the yield and adversely affecting the quality of the coffee. Detecting and controlling these diseases in their early stages represent formidable challenges, since traditional methods rely on visual observation by experts and often fail in accurate diagnosis. Machine learning (ML) techniques are alternative solutions for automating the classification of plant diseases. This research study aims to provide a comprehensive understanding of the strengths and weaknesses of various deep learning models and transfer learning approaches such as EfficientNetB0, MobileNetV2, CNN and VGG16, shedding light on their effectiveness in addressing the complexities of the multi-class label problem in the context of leaf disease detection in Arabica coffee leaves. Utilizing the “JMuBEN” dataset with 58,405 arabica coffee images across five classes consisting of four diseases such as Phoma, Cercospora, Leaf Rust, Miner and one class of healthy leaf images, the research aims to comprehensively assess each model's efficacy. The test accuracy of the investigated models ranged from 32.49% to 99.7%. EfficientNetB0 outperformed all the models with a test accuracy of 99.72% and an overall F1 score, Recall and Precision of 99.7%. Beyond Arabica coffee, the findings may extend to Robusta coffee and have broader applications in crop disease detection.