Deep learning technique for plant disease classification and pest detection and model explainability elevating agricultural sustainability.

The rapid advancement of technologies such as artificial intelligence (AI), deep learning, and precision agriculture tools is driving the development of efficient, data-driven crop management solutions. These innovations are increasingly critical in modern agriculture, where early and accurate detection of plant diseases plays a vital role in securing crop yields and sustainability. Agronomists, agriculturists, and local farmers continue to face significant economic losses due to delayed diagnosis or misclassification of diseases affecting high-value crops, key contributors to the global market. Failure to identify and manage such diseases in time can severely impact both agricultural productivity and global food supply chains. To achieve the United Nations’ sustainable development goals of zero hunger, climate change, good health, and well-being, early and timely disease detection is critical to ensure increased apple-related production, damage control, and reduced application of inappropriate herbicides that pollute the environment. Despite the availability of various methods for early disease detection and classification, how early signs of green attacks can be identified remains uncertain. Using the Turkey Plant Pests and Diseases (TPPD) dataset with 4,447 images categorized into 15 diverse classes, this research implements ResNet-9 to detect and classify the commonly known pests and diseases of six plants, including Malus pumila, Prunus armeniaca, Prunus padus, Prunus persica L. Batsch., Pyrus communis L., and Juglans regia. A laborious hyperparameter tuning, hyperparameter optimization, and augmentation procedure on the training set was done for some imbalanced dataset classes. Testing results of the proposed model demonstrated accuracy, precision, recall, and F1-score values of 97.4%, 96.4%, 97.09%, And 95.7%, respectively, which is a significant leap in comparison to other existing research. This study further elucidates and enhances the interpretability of the proposed model by making saliency maps available using SHapley Additive exPlanations (SHAP) that efficiently illustrate the rationale behind the model’s prediction capabilities. The study further tested the statistical significance of the model, the Area Under the receiver operating characteristic curve (AUC-ROC), and the confidence interval (CI). Critical observations revealed that the model uses several visual cues for disease detection and classification, including (i) edge contours and shape structures that help define lesion boundaries, (ii) texture and color variations that signal symptom type and severity, and (iii) high-activation regions that indicate areas of strong feature relevance. These cues collectively guide the model in distinguishing between visually similar disease patterns across different plant parts. The application of SHAP saliency maps further enabled interpretation by visually localizing and quantifying the influence of these features on the model’s predictions.