Interpretable Deep Learning for Sustainable Agriculture: CNN and LIME-Based Plant Disease Diagnosis

Abstract

The manual inspection forms the bulk of conventional plant disease identification methods, but it is inefficient, unreliable, and time-consuming for large-scale agricultural farms. Though they are promising, numerous automated systems that are currently employed struggle to be applied in practice due to their transparent decision-making processes and low capacity to adapt to real-world field conditions. This research places emphasis on readability in predictions while developing a clear and reliable deep learning model to diagnose diseases among 38 plant species. The objective of the research was to counteract the increasing threat of food insecurity, which is being exacerbated by population growth and climate change. Methods like picture augmentation and regularization ensure the model's adaptability to different environmental conditions. Explanation tools visually highlight specific aspects of disease-related images, which is crucial to adoption and boosts user trust. Agronomic information is also aligned with visual explanations. Through the reduction of unnecessary use of pesticides, the strategy not only supports sustainable agricultural practices but also propels global efforts in equitable agricultural innovation, poverty reduction, and climate resilience. This study presents a CNN-based model for multi-class plant disease classification across 38 categories using an augmented dataset of leaf images. The model achieved a high-test accuracy of 97.96% and demonstrated strong generalization with minimal overfitting. LIME-based explainability validated the model’s agronomic relevance, enhancing its trust and applicability in real-world agricultural diagnostics.