Search for a command to run...
This repository contains the data and machine learning code associated with the paper "Over 80% of Svalbard's cultural heritage faces retrogressive thaw slump exposure by 2100 under high-emissions scenarios". Ensemble susceptibility maps (ensemble_maps.zip) Pixel-wise mean RTS (Retrogressive Thaw Slump) susceptibility probability rasters at 100 m resolution for Svalbard, covering 29 scenarios: one historical baseline (2010–2025) and 28 future projections (4 SSP scenarios × 7 decadal periods from 2030–2040 to 2090–2100). Each raster represents the ensemble mean probability derived from 10 independent XGBoost model runs (seeds 42–51). CRS: EPSG:32633 (UTM Zone 33N). Cultural heritage susceptibility tables all_ch_point_susceptibility.xlsx: Mean RTS susceptibility values extracted at all 4,590 registered cultural heritage (CH) sites in Svalbard across all 29 scenarios, including 95% confidence intervals computed from the 10-run ensemble . priority_ch_point_susceptibility.xlsx: Same extraction for the 99 priority CH polygon sites, using zonal mean aggregation across each polygon footprint rather than centroid sampling. Model code train_xgboost_monotonic.py: Trains a monotonic-constrained XGBoost binary classifier for RTS susceptibility. Hyperparameters are selected via random search with 5-fold stratified cross-validation (100 trials). Monotonic constraints are applied to dynamic climate features (TDD +1, FDD −1, Tmax +1, Rainmax +1) to ensure physically consistent response directions. quantile_reclass.py: Computes global quantile bin boundaries from the union of all baseline and future climate rasters, then encodes dynamic climate features as ordinal bin indices. This encoding ensures that future out-of-distribution climate values are mapped within the training feature space, maintaining prediction reliability across climate scenarios. Model inputs include ERA5-Land derived climate factors (TDD, FDD, Tmax, Rainmax), CMIP6 ensemble projections bias-corrected via Quantile Delta Mapping (QDM), ArcticDEM-derived terrain factors (Slope, TWI), and SoilGrids fine-fraction estimates. Training labels are based on a Svalbard-wide RTS inventory mapped from optical and SAR imagery.