Explainable Ensemble Deep Learning Architecture for Fruit Disease Detection for Sustainable Agriculture and Food System Resilience

Abstract

Detecting diseases in fruits and vegetables is essential for reducing food waste and ensuring quality control across supply chains. Manual inspection is slow, inconsistent, and unscalable. To address this, we developed a deep learning framework that benchmarks five CNN-based architectures and integrates ensemble learning for robust fruit disease classification. We constructed voting-based ensembles using a dataset of 28 classes across 14 fruit and vegetable types to improve accuracy and stability. The best individual model, ResNet9, achieved 97.43%, while the best ensemble configuration (EfficientNetB0 + MobileNetV2 + ResNet9) achieved 98.47% test accuracy, outperforming all individual models and confirming the benefit of model combination for higher robustness and generalization. To enhance transparency, Grad-CAM was applied to visualize the model's focus on diseased regions, validating decision reliability. A unified framework diagram summarizes the end-to-end pipeline, from preprocessing to inference and explainability. The proposed system demonstrates strong potential for scalable, automated food quality inspection, contributing to sustainable agriculture and resilient food supply chains.