A Hybrid Lesion Segmentation and Deep Learning Framework for Intelligent Grape Leaf Disease Classification

Abstract

Early and accurate detection of grape leaf diseases is crucial for enhancing productivity and precision agriculture. But the traditional deep learning approach tends to classify a whole leaf image, and is easy to be interfered by the background and has poor performance under real-field conditions. The authors suggest a hybrid architecture combining improved SegNet-based lesion segmentation and an Attention-Guided GhostNet architecture with the proposed LGTrp feature learning strategy for classification of grape leaf diseases (AGh-ISegN-DL). The framework first localizes regions of interest in the image with the disease in order to block any irrelevant background information to include in the classification process, and then extracts the discriminative disease specific features for classification. Performance was tested with standard grape leaf disease datasets and also tested with cross-dataset validation with real field AgroVision data set (collected independently). The proposed framework performed better than the evaluated baseline models, such as the previously published Py-ICNN framework, with an accuracy of 96.07%, sensitivity of 96.08%, a precision of 96.05%, a specificity of 99.35% and an F1-score of 96.05%. Ablation analysis showed that the overall classification performance was improved by the synergistic effect of combining lesion segmentation and attention-guided feature learning. Across a new dataset of real-field images acquired independently, with natural environmental variability and class imbalance, the accuracy of 93.10% and the specificity of 97.60%, showed high robustness and generalisation capacity in practical conditions of vineyard. The proposed approach is overall efficient for grape leaf disease diagnosis with intelligence and it has strong potential use in precision agriculture and smart monitoring systems in the vineyard.