Towards a data-informed desktop study: Predictive modelling of geological and archaeological data

In early-stage infrastructure planning, understanding subsurface conditions and cultural heritage risks is essential to reduce costs, delays, and environmental impact. We present a comparative approach to predicting lithological complexity and archaeological risk along a proposed railway corridor in southern Sweden. Using publicly available datasets, a geological classifier was trained on 24,000+ well records to predict lithological heterogeneity, and archaeological models were developed for selected site categories using curated records. We evaluate gradient-boosting and related learners (CatBoost, XGBoost, LightGBM, Random Forest) under nested spatial cross-validation with position-aware features that encode geographic structure, and we contrast these results with Maximum Entropy models for archaeology. Under spatial cluster CV, the more realistic evaluation, model differences for geology are small; CatBoost is marginally highest, and all models exhibit the expected drop from non-spatial to spatial validation. The archaeological models achieve good discrimination but display sensitivity to sampling bias and geographic transfer, reflecting differences in data curation and coverage between domains. The best-performing models are used to generate continuous maps that can be combined with planning suitability layers to support early route selection, fieldwork prioritisation, and risk-informed decisions. Overall, model utility is governed less by algorithm choice than by data standardisation, labelling quality, and explicit handling of spatial autocorrelation. Future work should unify geological and archaeological predictors in multi-criteria frameworks and continue improving the availability and quality of national subsurface datasets.”