A 1-km monthly surface air temperature dataset for Arctic land (2000–2025)

Monthly mean near-surface (2 m) air temperature at 1-km resolution over Arctic land for March 2000 to December 2025 (310 monthly fields), reconstructed by an ensemble of Random Forest, LightGBM and XGBoost regressors trained separately for each calendar month on station observations (CRUTEM5 station archive, release 5.1.0.0) with MODIS predictors: Terra and Aqua day-time and night-time land surface temperature (MOD11A2, MYD11A2), Terra and Aqua NDVI (MOD13A3, MYD13A3), MCD12Q1.061 LC_Type5 land-cover fractions, latitude and longitude. Elevation is deliberately not a predictor; the reasons, and the measured consequence, are given in the accompanying paper. What changed in version 3.0. The predictor values of the training table are now pooled from the same gridded stacks the prediction reads, rather than from a separate extraction of the granules, so that training and prediction use one source. The land surface temperatures entering training are screened with three checks of the GHCN-Daily quality-assurance system at their published thresholds (world-record exceedance, a climatological outlier check at six biweight standard deviations, and spatial corroboration against the nearest neighbours); 3860 values, 0.50 % of the land surface temperatures, were found to be invalid — seven satellite month-year composites had reached the table with MODIS fill values that had not been masked when the monthly means were formed — and are treated as missing. The gridded predictor stacks were checked for the same signature and carry none of it, so no field of version 2.0 was affected by that defect. The models were refitted and the whole record re-predicted. Hold-out accuracy at the 120 withheld stations improves from 1.50 to 1.48 °C root-mean-square error. Data are on the native MODIS 1-km sinusoidal grid (34 tiles h09–h26, v00–v02, mosaicked to 3600 × 21600 cells, pixel 926.625433 m) with 2-D latitude/longitude and a CF grid mapping, so no resampling has been applied to the values. Water pixels are removed with the MOD44W mask and Greenland is excluded. Each annual file (ArcticSAT1km_monthly_YYYY.nc, CF-1.8 netCDF4) contains: sat — monthly mean near-surface air temperature (°C), ensemble mean of the three models; sat_model_spread — the per-pixel standard deviation of the three individual model predictions for the same month, a measure of structural (model-choice) uncertainty. Both are stored as int16 with scale_factor 0.01 °C (precision 0.01 °C), zlib-compressed, chunked one MODIS tile per chunk. ArcticSAT1km_grid.nc holds the static 2-D lat, lon, the MOD44W land_mask and tile_index; ArcticSAT1km_climatology.nc holds the monthly climatologies, the warm (April–October) and cold (November–March) seasonal means, the number of valid years per pixel and the mean inter-model spread. Note on coverage: a fixed subset of land cells never has a valid MODIS/Terra night-time LST retrieval. The pattern is static — identical in every month and year of a given tile — and varies strongly between tiles (measured examples, as a fraction of the tile's land cells: h22v02 0 %, h24v02 12.3 %, h19v02 43.4 %). There the missing night-LST predictor enters the models as zero, so the prediction rests on the remaining predictors and is less well constrained. Nothing is blanked and no quality-flag variable is provided; every file carries this note in its comment global attribute, and README_data.md lists it under "Known limitations". The climatologies in ArcticSAT1km_climatology.nc use the period 2001–2025 (25 years), the earliest period in which all twelve calendar months are complete, since January and February begin in 2001. MANIFEST_md5.txt allows md5sum -c verification, MANIFEST_sizes.tsv lists file sizes, PROVENANCE.json gives a machine-readable provenance record, and README_data.md documents every variable, the method and the known limitations. Processing code: https://doi.org/10.5281/zenodo.23490034 (https://github.com/SeoyeonLee-rosee/arctic-1km-sat).