Optimized Machine Learning Framework for Detection of Banana Leaf Disease

Abstract

Introduction: Banana plants are prone to various diseases which significantly impact yield and agricultural productivity. Leaf diseases of banana such as Cordana and Sigatoka significantly affect global banana production, making early and accurate detection essential for crop health management. This research proposes an optimized approach based on machine learning for automatic detection and classification of banana leaf disease. The methodology includes image pre-processing, augmentation, RGB color to L*a*b*color space conversion, K-Means segmentation followed by thresholding. This is followed by feature extraction (GLCM, LBP, Hu moments, Color features), PCA-based dimensionality reduction, and finally classification using SVM, KNN, Random Forest, and Gradient Boosted Stacking Ensemble models. The results obtained demonstrate the efficacy of the features after Principal Component Analysis (PCA) reduction, with the Gradient Boosted Stacking Ensemble model achieving the highest accuracy among the different models. The best classification performance was obtained with the Gradient Boosted Stacking Ensemble which achieved an accuracy of 95.53%. The proposed model outperformed other individual models like Random Forest, KNN and SVM which achieved accuracies of 78.45%, 84.71% and 82.22% respectively. Further, performance evaluation using precision, recall, F1-score, ROC curves, AUC and confusion matrices validates the robustness of the proposed ensemble method in classifying healthy and diseased banana leaves. The framework thus developed provides a scalable and reliable solution for automated detection of banana leaf disease which will support early disease management strategies in precision agriculture.