Enhancing interpretability and fidelity in convolutional neural networks through domain-informed knowledge integration

Abstract

This study addresses the need for robust disease detection methods in vegetable crops by introducing a novel initialization method for convolutional neural networks (CNNs). Rather than creating a new CNN architecture, our approach focuses on infusing expert knowledge from phytopathology directly into the model’s foundation. This innovative initialization ensures that the CNN possesses a contextual understanding of intricate disease patterns specific to tomatoes. Additionally, our study redefines the role of heatmaps as a dynamic metric for assessing model fidelity in real-time. Unlike traditional post hoc applications, heatmaps are integrated into the model evaluation process, providing insights into decision-making processes and alignment with expert-derived expectations. This dual innovation aims to enhance transparency and fidelity in CNNs, offering a nuanced and effective solution for disease detection in agriculture. The study contributes to advancing artificial intelligence applications in agriculture by providing accurate predictions and a deeper understanding of the underlying decision mechanisms crucial for crop health management.