Explicit Geological Constraints + Data‑Driven Modeling Improves Cross‑Regional Mineral Prospectivity and Interpretability

Background and Problem

Mineral prospectivity mapping (MPM) is essential for reducing exploration risk, yet ore‑forming processes are governed by multiple geological factors that create strong spatial non‑stationarity and anisotropy. Existing machine‑learning and graph‑based models typically handle these spatial characteristics implicitly, limiting geological interpretability and prediction stability.

Proposed Method

The team proposes an anisotropic convolution attention‑weighted logistic regression framework (ACAWLR). First, directional weighted covariance analysis extracts primary and secondary ore‑forming directions, defining an anisotropic spatial distance metric. An anisotropic spatial proximity neural network (ASPNN) learns direction‑dependent spatial relationships, which are then embedded into a convolutional attention network coupled with a logistic regression model. This explicit incorporation of anisotropic spatial adjacency quantifies heterogeneity and directionality of ore formation while preserving predictive accuracy.

Datasets

Meguma terrane (Canada) gold dataset : a classic benchmark with comprehensive deposit data and clear ore‑forming background, used for fine‑scale evaluation.

Southern Cordillera (USA) porphyry copper dataset : a large, multi‑state region with complex tectonic and magmatic activity, used to test cross‑regional scalability.

Experimental Setup and Model Comparison

Multi‑scale, hierarchical validation was performed. In the Canadian benchmark, ACAWLR was compared against geographic weighted logistic regression (GWLR), support vector machine (SVM), random forest (RF), multilayer perceptron (MLP), and graph attention network (GAT). The same model was later applied to the US copper province to assess robustness in a complex, large‑scale setting.

Results

ACAWLR achieved the highest recall and overall predictive performance among all tested methods. It balanced accuracy, spatial generalization, and geological interpretability, producing continuous prospectivity maps that align with geological knowledge and fully recover known deposit locations.

Interpretability Analysis

Region‑level analysis revealed distinct controlling factors. In the porphyry copper system, copper content emerged as the dominant predictor, while lithology, faulting, and gravity anomalies showed varying spatial influence across different tectonic contexts. Anisotropic analysis identified ore‑forming directions consistent with regional structural regimes, offering intuitive guidance for exploration planning.

Additional Team Contributions

The Zhejiang University Earth Science team has also released a series of open‑source geospatial AI models. Their GNNWR package (≈50 000 downloads and citations) addresses spatial‑temporal non‑stationarity, while the attention‑based CatGWR model incorporates scenario similarity into geographic weighted regression. Furthermore, the heterogeneous contrastive graph fusion network (HCGFN) enables joint classification of hyperspectral and LiDAR data. Relevant papers are cited with URLs in the source.

Conclusion

By explicitly embedding geological constraints through anisotropic spatial proximity and attention mechanisms, the proposed ACAWLR framework significantly enhances both the performance and interpretability of cross‑regional mineral prospectivity predictions, providing a new technical pathway for data‑driven exploration.