Remote Sensing-Enhanced Geologically Constrained Reinforcement Learning with Dynamic Spatial Search for Mineral Prospectivity Mapping

Mineral prospectivity mapping typically relies on static evaluations of individual research units. Although existing spatial modeling methods capture spatial patterns and relationships, they rarely represent spatial relationships as a sequential process of decision making that depends on the current state. This limits their ability to represent spatial changes and cumulative effects. This study proposes a geologically constrained reinforcement learning framework with dynamic spatial search for mineral prospectivity mapping. The method uses a graph neural network to identify geochemical anomalies. It also integrates favorable geological conditions with surface spectral information to construct a geological environment score. Remote sensing responses further refine the spatial representation and differentiation of favorable geological environments. These scores define an integrated prospectivity field and a local geological constraint that guide an agent to move among neighboring units or stop searching. During inference, each research unit serves as an independent starting state, and its prospectivity is evaluated using both the stop-action value and path-averaged prospectivity generated by the learned policy. The chromite case study in Balikun, Xinjiang, indicates that the proposed method identifies high prospectivity areas showing some spatial correspondence with known deposits and exhibits relatively good predictive performance compared with the selected baseline methods. A comparison with a model that excludes remote sensing information provides case-specific evidence of the contribution of these enhanced geological constraints to prospectivity prioritization when the exploration area is fixed. Ablation analysis and other experiments further support the effectiveness and validity of the proposed method.