A Heuristic Approach to Ensemble-Based Deep Learning Models for Plant Disease Classification and Farming Decision Support

Abstract

Leaf based plant diseases detection is one of the significant factors affecting crop yield and productivity. Most of the current techniques used for disease prediction are trained using observational records featuring numerous plant image parameters, with a higher frequency of diseased images compared to blight-free images. Hence, discriminating against the crucial insights from irrelevant and redundant images has been a crucial and challenging study. This research inspects the suitability of machine learning models in disease prediction focusing both specific and wide range of plant leaf images. Also, most classical methods are pretentious by various issues such as the format of image statistics, computation, and representation. To address this crucial setback in the present prediction methods, the proposed system develops a hybrid model utilizing stacked ensemble learning, which enhances the detection of plant disease attacks beyond what conventional learning methods. The proposed stacked ensemble-based disease prediction framework is designed to identify both misclassified and correctly classified images. This approach features a two-tier classification mechanism that involves a base learner (Level 0) and a meta learner (Level 1). It considers both image datasets and image features as inputs to facilitate the two-tier classification process. It also focuses on extracting internal features from the damaged leaves. The proposed model was trained with over 30,000 images at various levels. The experimental results revealed that the stacked ensemble learning technique outperformed with a prediction accuracy of 99.93%.