GeoXAttn-Net: Explainable Cross-Attention Network for Ore–Waste Block Classification using Geological and Operational Mining Data

Abstract

Recently, artificial intelligence and deep learning methods have received significant attention in miningapplications because of their ability to support mineral resource evaluation, production planning, and ore-waste decisionmaking. However, ore-waste block classification in open-pit mining environments remains challenging due to complexgeological variations, spatial heterogeneity, economic uncertainty, and the dependency between geological and operationalmining parameters. This study develops GeoXAttn-Net, an Explainable Cross-Attention Network for ore-waste blockclassification using geological and operational mining data. Following preprocessing, feature engineering computes profitmargin, ore-to-waste economic ratio, and spatial depth-related information. SHAP-guided dynamic feature weightingadaptively refines feature contributions, while separate geological and operational encoders learn physical and economiccharacteristics. Bidirectional cross-attention captures dependencies between these representations, and transformer-basedfusion with SHAP-guided attention refinement improves feature interaction learning and interpretability. A denseclassification head with softmax activation performs final ore-waste classification. Evaluation on the Open-Pit MiningBlock Model Dataset demonstrates improved classification performance and interpretability, supporting intelligent miningdecision-making, mineral resource evaluation, and economically sustainable mine planning.