Dataset and code for: "Climate-Constrained Learning Reveals Rising Retrogressive Thaw Slump Susceptibility Across Northern Hemisphere Permafrost"

This repository contains the data and machine learning code associated with the paper "Climate-Constrained Learning Reveals Rising Retrogressive Thaw Slump Susceptibility Across Northern Hemisphere Permafrost".Ensemble susceptibility maps Pixel-wise RTS (Retrogressive Thaw Slump) susceptibility rasters at 1 km resolution across the Northern Hemisphere permafrost domain, covering 29 scenarios: one baseline (2010–2025) and 28 future projections (4 SSP scenarios × 7 projection windows from 2030–2040 to 2090–2100). For each scenario, three rasters are provided:• probability/: ensemble mean susceptibility probability (continuous 0–1), averaged across 10 independent PAMR-Net runs.• class/: five equal-width susceptibility classes (Very Low to Very High), obtained by classifying the ensemble mean probability at thresholds [0.2, 0.4, 0.6, 0.8].• std_probability/: run-to-run standard deviation of the susceptibility probability across the 10 runs.CRS: EPSG:3995 (Arctic Polar Stereographic). Resolution: 1000 m. All rasters are masked to the mean annual ground temperature (MAGT) < 0 °C permafrost domain (Obu et al., 2019). NoData: −9999 for probability and standard deviation rasters, 0 for class rasters.Model code• train_pamr_net.py: Trains PAMR-Net (Physics-Anchored Monotone Response Network), a decomposed-logit neural network for RTS susceptibility. The model separates a static-site baseline term from a constrained climate-response term, with monotone, sign-consistent responses to the four climate factors (increasing TDD, Tmax, Rainmax, and decreasing FDD can only maintain or increase susceptibility). Non-negative weights are enforced through softplus reparameterization. Hyperparameters are selected via Optuna with a TPE sampler (40 trials), and the model is trained over 10 repeated runs (seeds 42–51) to form an ensemble.• predict_raster.py: Applies a trained PAMR-Net ensemble to the baseline and bias-adjusted CMIP6 climate fields, producing the susceptibility rasters for each SSP and projection window.• logit_decomposition.py: Decomposes the projected logit change into a static-site baseline, a baseline climate response, and per-factor future climate increments, supporting the attribution analysis in the paper.Model inputs include ERA5-Land derived climate factors (TDD, FDD, Tmax, Rainmax), NASA NEX-GDDP-CMIP6 projections bias-corrected via Quantile Delta Mapping (QDM), MERIT DEM-derived terrain factors (slope, TWI), and SoilGrids 2.0 fine-fraction estimates. Training labels are based on a Northern Hemisphere RTS inventory compiled from published datasets (see Table S1 of the paper). The original RTS inventories are not redistributed in this repository and are available from their respective sources listed in the paper.A small synthetic demo dataset is included for code testing only; it does not contain real RTS locations.