Dataset and code for: "Over 80% of Svalbard's cultural heritage faces retrogressive thaw slump exposure by 2100 under high-emissions scenarios"

Read full text Summarise with VIRAT
Authors: Ionuţ Cristi Nicu, Hualiang Zhu, Xianwei Zhang, Luigi Lombardo
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

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.

Fields of Study & Subjects

Climate change and permafrost Indigenous Studies and Ecology Cryospheric studies and observations

Publication details

Journal / source
University of Twente Research Information
Publication date
23 March 2026
Language
EN
Document type
Dataset