Optimizing Soursop leaf disease classification with a lightweight ensemble model and explainable AI

Abstract

Traditional deep-learning methods to detect plant leaf disease can be complex and time-consuming if image numbers and size increase. Moreover, complex deep learning networks take longer and require larger memory to produce results. However, feature extraction methods provide some advantages in such a scenario. Using heavy-weighted models to enhance accuracy without considering the long execution time is a drawback of research. A weighted model increases the time and space complexity of an experiment. Considering the mentioned limitations, this study proposes a lightweight model experimenting with six deep feature extraction models, five feature selection models, and four machine learning classifiers. During the experiment, a soft voting ensemble classifier was developed to remove a single classifier's limitations and the unstable performance of the standalone classifiers. After a rigorous experiment, the (ResNet101 – RFE – Ensemble Classifier) together formed the best performer Soursop Ensemble (S-Ensemble) model that obtained a test accuracy of 99.6 % with an execution time of 648.05 s, outperforming other models. The whole experimental analysis was performed on a primary Soursop leaf disease dataset with six classes containing 3838 images. Finally, the Explainable AI (XAI) model Local Interpretable Model-agnostic Explanations (LIME) is used to interpret the reasons behind the best-performer and lowest-performer models' performance. LIME visually highlights which leaf regions influence each prediction, helping users understand model behaviour and enhancing its practical usability in real-world agricultural settings. This research aims to assist farmers with detecting Soursop leaf disease with less execution time and offer researchers an in-depth preview of deep feature-based detection and classification technology to detect and classify diseases within a short training time.