Interpreting potato disease classification using explainable artificial intelligence

Abstract

Timely and accurate crop diseases detection is most important for ensuring global food security. For detecting diseases in crops, many different machine learning (ML) models were proposed. These models work as a black-box, and without proper explanation of these models’ decisions, farmers may find it difficult to trust these systems. For this, many model explainability methods were also proposed. All these methods have been evaluated qualitatively, but their quantitative and cross-comparison are missing. This study addresses this gap by emphasizing quantitative validation of models’ explanations. This study investigates the interpretability and performance of two approaches for potato leaf disease classification: i) manual feature engineering with relief-based feature selection and artificial neural network (ANN) classification with local interpretable model-agnostic explanations (LIME) and ii) deep convolutional neural networks (CNNs) based on mobile network version 2 (MobileNetV2) with gradient-weighted class activation mapping (Grad-CAM). Quantitative assessment of explanation quality was performed using fidelity and robustness metrics. While LIME achieved a lower average fidelity drop 10.90% and higher robustness 84.27% compared to Grad-CAM 18.68% fidelity drops and 79.38% robustness, Grad-CAM provided clearer visual explanations. These findings suggest that explanation effectiveness is influenced by model design and data characteristics, highlighting the need for careful selection of interpretability techniques in precision agriculture applications.