A Robust Machine Learning Framework for Banana Damage Classification Using Image Processing Techniques for Smart Agricultural Quality Assessment

Abstract

Reliable banana quality assessment remains challenging due to the subjectivity of manual inspection. This study proposes a robust and evaluation-driven framework for banana damage classification that integrates image preprocessing, texture-based feature extraction, and multi-model learning strategies. A dataset of 568 banana images categorized into three damage levels was systematically processed to extract discriminative texture descriptors, reflecting surface variations and bruising patterns observed in practical conditions. To ensure statistical reliability, multiple training–validation–testing configurations were employed to evaluate model stability across varying data distributions. Five machine learning classifiers, namely Updateable Multiclass Classifier, Sequential Minimal Optimization (SMO), Multilayer Perceptron (MLP), Random Forest, and Iterative Classifier Optimizer, were evaluated to investigate the relationship between texture representation and classifier behavior. Comparative results demonstrate that adaptive classification models achieve superior performance, with the best model reaching 91.11% accuracy and an F1-score of 0.91, while maintaining consistent performance across experimental settings. Performance depends more on feature–model compatibility than model complexity. This work contributes a reproducible and practically applicable framework for intelligent quality assessment, supporting scalable deployment in smart agriculture and reducing post-harvest losses through data-driven decision-making. The proposed framework can support automated banana grading systems in post-harvest operations, reducing inspection subjectivity and enabling scalable deployment in smart agriculture environments.