Multi-Label Based Plant Disease Identification and Treatment Recommendations Using Transfer Learning and Attention Mechanism

Abstract

Agriculture remains the primary contributor to the economy and the foundation for food security in Ethiopia, where small farmers grow crops such as enset, tomato, and pepper. However, plant diseases are one of the most pressing concerns for farmers, particularly in the Wolaita Zone, resulting in a reduction in crop production levels and the feasibility of their treatment. To mitigate this problem, previous studies have conducted research on plant disease classification. However, there is a lack of studies conducted to classify plant diseases and provide treatment recommendations. Thus, the proposed study employed a design science research design with both qualitative and quantitative approaches. To achieve our objective and fill the research gap, we developed a lightweight multilabel classification model for crop disease detection and treatment recommendation using transfer learning with pretrained convolutional neural networks, including YOLOv8, VGGNet, DenseNet201, and MobileNetV2 integrated with an attention mechanism that extracts features used for model training. A combined 12,490 images dataset was prepared from healthy and infected enset, tomato, and pepper leaves, as well as the CSV file containing the disease types and their treatment recommendations. Experiments for the proposed models are carried out, and the results of the DenseNet201 algorithm fused with the Squeeze-and-Excitation (SE) attention mechanism resulted in a Hamming Loss of 0.0197, Validation Accuracy of 0.992, and mAP of 0.9554 over the entire class. MobileNetv2 demonstrated the best performance in terms of computational complexity, making it ideal for device-level execution.