A Comparative Analysis of Multi-Class and Binary Papaya Leaf Classification Using Optimized Darknet Architectures

Abstract

Accurately determining the health status of papaya leaves is of great importance for early detection of diseases and the continuity of sustainable agricultural production. In this study, the performance of two improved Darknet architectures, Darknet-19 and Darknet-53, was evaluated on both the five-class disease recognition problem and the Healthy–Unhealthy binary classification scenario. The used BDPapayaLeaf dataset contains a total of 2,159 highresolution leaf images belonging to five categories: Anthracnose, Bacterial Spot, Curl, Ring Spot and Healthy. Various training methods such as data augmentation, label smoothing, dropout and cosine-based learning rate planning were applied to increase the generalization ability of the models. Looking at the five-class classification results, while the Darknet-19 model achieved 79.81% accuracy, the Darknet-53 model with a deeper structure reached 84.78% accuracy. This result shows that residual layers provide a significant advantage especially in capturing complex disease patterns. When binary classification is performed, the accuracy rates increased significantly; Darknet-19 achieved 97.20% accuracy, while Darknet-53 achieved 98.14% accuracy. The findings demonstrate that optimized Darknet architectures offer a reliable, effective, and applicable approach to automatically identifying papaya leaf diseases.