Predictor stack, terrace densities and model outputs for: Spatiotemporal turnover of the drivers of terracing on the Loess Plateau, 2000-2020

Agricultural terraces are the widest soil and water conservation measure on China's Loess Plateau, yet the drivers of the terrace pattern that already exists remain poorly served. This deposit contains the data and code behind that analysis: eleven open predictors linked to three-period (2000, 2010, 2020) 30 m terrace density on a 0.01 degree grid (660,997 cells), scored with OLS, GWR, GNNWR and GTNNWR under latitude-band holdouts run in both directions, plus tree ensembles and a pooled panel. The deposit is organised by the argument of the paper: 00_SSOT lists every number in the manuscript; 01_holdout_transfer holds the out-of-extent results, including GNNWR over five fixed seeds for both bands and three unseeded refits; 02_driver_analysis holds Tables 1-3 with cell-resampling intervals; 03_ablation_baseline, 04_robustness and 05_coefficient_fields hold the ablation, the robustness checks and the coefficient fields behind Figures 4-10; 06_code holds the twelve analysis scripts. GNNWR is reported as a median with its range over five fixed seeds, because run-to-run spread on the sparsely terraced northern band is 6 to 41 times the southern one. The earlier value of 0.323 for the southern 2010 band was not reproducible and has been withdrawn. See README.md before re-running anything.