Data and code for : Short-term meteorological drought predictability across Northern China: Local hydroclimatic information and a two-month skill horizon under partial target-window observation

This repository contains the code and supporting data products for a machine-learning study of short-lead meteorological drought predictability across Northern China. The analysis predicts the 3-month Standardized Precipitation Evapotranspiration Index (SPEI03) at leads of 1–6 months and evaluates the relative contributions of recent local hydroclimatic conditions, antecedent drought, and large-scale climate indices. The deposited code covers: - ERA5 and ERA5-derived drought-index preprocessing;- climate-index parsing, quality control, and standardization;- construction and quality control of the Northern China target grid;- XGBoost and random-forest model training;- persistence and linear autoregressive baselines;- training-period SHAP-based predictor selection;- model-output merging and analysis;- predictor-compression assessment;- temporal-robustness and annual-block bootstrap analyses;- reproduction of the manuscript figures. The deposited data products include: - the exact ERA5 grid-point lists defining the study domain and training sample;- English climate-index source and quality-control metadata;- the NOAA monthly climate-index compilation used in preprocessing;- an ETOPO 2022 elevation subset used only as the cartographic background in Figure 1;- training-SHAP feature-selection records for the 36-, 24-, and 18-predictor models;- point-level and Northern-China-aggregated model-skill tables;- aggregated feature- and category-level SHAP summaries;- predictor-compression and uncertainty-analysis results;- merged prediction records for the full and local XGBoost models and drought-memory baselines over the 1996–2005, 2006–2015, and 2016–2025 test periods. The archive contains 9,684 geometry-selected ERA5 grid points, of which 9,683 satisfy the target-availability criterion used for model training. One offshore grid point was excluded because SPEI03 was missing throughout the earliest training-period quality-control interval. Original ERA5 NetCDF files, reconstructed Zarr stores, and standardized ERA5 predictor stores are not redistributed because they are large and can be reconstructed from the cited Copernicus Climate Data Store products using the deposited code. Complete China Meteorological Administration National Climate Center monthly index values are also not redistributed because explicit third-party redistribution terms were not documented. The AMap-derived administrative-boundary shapefile is omitted for the same reason; the exact analyzed grid is instead preserved in the deposited target-point tables. The principal modeling environment used Python 3.11.13, XGBoost 3.0.2, SHAP 0.48.0, and scikit-learn 1.7.0. Machine-specific paths have been removed from the public code. Detailed file descriptions, provenance, known source-coverage limitations, reconstruction instructions, and SHA-256 checksums are provided in `README_data.md`, the code `README.md`, and the accompanying file-specific documentation. All deposited CSV files use UTF-8 encoding and English field names. Third-party datasets remain subject to their respective source licences and terms of use.