Addressing Background Bias in Explainable Orange Fruit Disease Classification Using Deep Learning

Fruit diseases significantly impact agricultural productivity, yet automated detection systems often fail to provide interpretable predictions and are sensitive to background variations in images, particularly in orange fruit disease datasets. Current deep learning approaches are prone to background bias, which reduces explainability and generalization. To address this, we propose a deep learning framework that explicitly reduces background noise and bias in orange fruit disease image classification while providing interpretable, pixel-level predictions. The framework integrates existing architectural components, including grouped convolutions with channel shuffling, Leaky ReLU and clipped ReLU activations, and attention-based feature extraction, within a bias-aware design motivated by explainability analysis. The contribution lies in the problem-driven integration of these components and a background standardization preprocessing step to improve explanation reliability. A Grid Search algorithm is used to optimize the hyperparameters. Data augmentation is applied to enhance generalization. We used perceptual hashing to ensure no duplicate images existed between training and testing sets, thereby preventing data leakage and maintaining dataset integrity. For interpretability, we employ Local Interpretable Model-agnostic Explanations (LIME); however, initial explanations highlighted irrelevant background regions. To address this, we introduce a novel preprocessing step using the GrabCut algorithm and morphological operations to standardize image backgrounds, ensuring explanations focus solely on diseased regions. Unlike existing methods, our background standardization technique, based on GrabCut and white background standardization, improves the relevance of LIME explanations by reducing background-focused attributions from 62.2% to 7.7% of cases, while yielding modest, consistent improvements in classification accuracy (0.15%–0.24%). We further evaluate DeepOrangeNet’s feature extraction by classifying its learned representations using six classifiers, including linear discriminant analysis, fine decision tree, Gaussian Naive Bayes, fine k-nearest neighbors, linear support vector machine, and logistic regression, demonstrating its superior adaptability. DeepOrangeNet has been compared with the state-of-the-art methods, proving not only its accuracy but also its explainable and lightweight architecture for real-world agricultural implementation.