Multi-Source Data Fusion and AutoGluon-Tabular Learning for Accurate Open-Pit Mine Boundary Delineation

Abstract

This study proposes an AutoGluon-Tabular workflow for open-pit mine boundary delineation from multi-source remote-sensing mosaics. The method converts optical, SAR, terrain, texture and contextual mosaic descriptors into object-level records, uses automated stacked tabular learning to estimate boundary probability, and reconstructs vector boundaries through calibrated probability gradients. The workflow is designed for mining scenes where active pits, waste dumps, haul roads, bare soil and shadowed highwalls have overlapping spectral responses, making boundary extraction different from ordinary mine-area classification. By combining object-level predictors with uncertainty-aware vector cleaning, the approach aims to reduce fragmented boundaries and focus manual review on ambiguous segments. Experiments on 18 mining districts and 42,600 labelled objects show an F1-score of 0.924, an IoU of 0.873, a boundary F1-score of 0.846 within a 20 m tolerance, and a mean boundary offset of 7.9 m. Compared with random forest, XGBoost, LightGBM-only and compact U-Net baselines, the proposed workflow improves geometric boundary quality while retaining interpretable feature groups and manageable computational cost.