Research on Crop Leaf Disease Identification and Severity Assessment Based on Lightweight Multitask Deep Networks

Abstract

In response to model 1, we first cleaned and standardized 61 crop disease image categories by removing duplicates through comparing image filenames with label files using provided path information. Valid samples were resized and augmented to construct a multi-disease classification model based on the lightweight MobileNetV3-Large, with category IDs mapped to disease names. The model was trained and validated with cross-entropy loss, AdamW optimizer, and cosine annealing learning rate, with epoch-dependent loss and accuracy curves recorded. For model 2, a few-shot recognition solution was developed based on model 1, retaining 10 training samples per category. Using pre-trained MobileNetV3-Large as the feature backbone (parameters < 20M), only upper convolutional and classification layers were fine-tuned. Enhanced augmentation, label smoothing, and cosine annealing mitigated overfitting and class imbalance, achieving ~73% validation accuracy for 61 categories; Grad-CAM confirmed the model focuses on leaf lesions. Regarding model 3, severity-graded prediction was implemented by mapping 61 diseases to 3 severity levels via appendix JSON annotations and disease description tables. Images were regrouped to build a three-classification dataset, and a severity prediction model with MobileNetV3-Large (transfer learning, augmentation) was trained, outputting overall accuracy, macro-F1, recall, and a confusion matrix; Grad-CAM visualized key lesions for high-confidence correct predictions. For model 4, a lightweight integrated multi-task model was developed for simultaneous disease identification and severity assessment, using MobileNetV3-Large as the shared feature backbone with 61-category disease and 3-category severity classification heads. Joint optimization via multi-task loss enabled feature sharing and fine-grained assessment, with joint accuracy, confusion matrices, and Grad-CAM analyzing synergy and lesion focus, supporting interpretable diagnosis reports.