Model-agnostic local explanation: Multi-objective genetic algorithm explainer

Late detection of plant diseases leads to irreparable losses for farmers, threatening global food security, economic stability, and environmental sustainability. This research introduces the Multi-Objective Genetic Algorithm Explainer (MOGAE), a novel model-agnostic local explainer for image data aimed at the early detection of citrus diseases. MOGAE enhances eXplainable Artificial Intelligence (XAI) by leveraging the Non-dominated Sorting Genetic Algorithm II (NSGA-II) with an adaptive Bit Flip Mutation (BFM) incorporating densify and sparsify operators to adjust superpixel granularity automatically. This innovative approach simplifies the explanation process by eliminating several critical hyperparameters required by traditional methods like Local Interpretable Model-Agnostic Explanations (LIME). To develop the citrus disease classification model, we preprocess the leaf dataset through stratified data splitting, oversampling, and augmentation techniques, then fine-tuning a pre-trained Residual Network 50 layers (ResNet50) model. MOGAE’s effectiveness is demonstrated through comparative analyses with the Ensemble-based Genetic Algorithm Explainer (EGAE) and LIME, showing superior accuracy and interpretability using criteria such as numeric accuracy of explanation and Number of Function Evaluations (NFE). We assess accuracy both intuitively and numerically by measuring the Euclidean distance between expert-provided explanations and those generated by the explainer. The appendix also includes an extensive evaluation of MOGAE on the melanoma dataset, highlighting its versatility and robustness in other domains. The related implementation code for the fine-tuned ResNet50 and MOGAE is available at https://github.com/KhaosResearch/Plant-disease-explanation . • MOGAE eliminates several critical hyperparameters LIME requires, simplifying the explanation process. • MOGAE delivers more accurate and interpretable explanations than LIME, enhancing the reliability of the model’s decisions. • The integration of adaptive Bit Flip Mutation with densify and sparsify operators in NSGA-II significantly improves the algorithm’s exploration capabilities within MOGAE. • Comprehensive evaluations of MOGAE are conducted using both the citrus leaf diseases dataset and the melanoma dataset, demonstrating its versatility and effectiveness across different domains.