A Computational Approach To Banana Leaf Disease Detection Thru Classic Image-Filtering Algorithms

Abstract

Banana crops in tropical regions are increasingly threatened by fungal leaf diseases such as Cordana, Sigatoka, and Pestalotiopsis. Traditional computer vision approaches for plant pathology often rely on edge-based filtering, which struggles under variable illumination and noisy textures. This study presents a revitalized classic approach that leverages multi-scale Gabor filtering, CLAHE-enhanced contrast, and adaptive Otsu thresholding to segment symptomatic regions from both ground and UAV-captured images. Haralick texture and color moment descriptors are extracted and classified using a Random Forest ensemble. Performance was benchmarked against SVM-RBF and MLP baselines under a leaf-exclusive 5-fold cross-validation scheme. The proposed method achieved 94.2 % accuracy and a macro F1-score of 0.942, outperforming both baselines significantly (p < 0.05). The pipeline maintains interpretability and low latency (0.038 s/tile), making it suitable for integration into lightweight agricultural drones and diagnostic tools for smallholder farmers.