HiGNN_LSTM: Dataset for training and evaluating TWSA forecasts

This dataset contains three preprocessed NetCDF files used to train, evaluate, and benchmark the HiGNN-LSTM and ConvLSTM2D models for global Terrestrial Water Storage Anomaly (TWSA) forecasting, as described in [your paper citation]. All files are on a global 1° × 1° regular grid with monthly temporal resolution and longitudes in the −180°–180° convention. The processing pipelines for these datasets can be found at https://github.com/viola1593/HiGNN-LSTM/tree/main/scripts/preprocess_data. TWSA-REC_era5_detrend_latavg.nc — Training and test dataset (1979–2023) Used to train and evaluate both models. Contains reconstructed TWSA from Li et al. (2021) [1] combined with ERA5 monthly mean climate fields [2]: sea surface temperature, surface pressure, 2 m air temperature, 10 m wind speed, total precipitation, evaporation, potential evaporation, runoff, volumetric soil water layers 1–4, leaf area index (high and low vegetation), and land-sea mask. All time-varying variables are linearly detrended at each grid cell with scipy.signal.detrend. Both sources are coarsened to 1° using spherical area-weighted averaging. GRACE-FO_era5_detrend_1deg_latavg_tempmatch.nc — Evaluation dataset (2018–2025) Used to evaluate model predictions against observed GRACE/GRACE-FO data. Contains TWSA from the CSR RL06.03 mascon product [3, 4] combined with the same ERA5 variables as above, processed identically. GRACE observations are temporally matched to ERA5 months by nearest midpoint within a ±20-day tolerance; original GRACE timestamps are retained in the GRACE_time coordinate for traceability. Li_csr_fcast/combined.nc — Reference forecast for comparison Used to benchmark model predictions against an independent observation-driven TWSA forecast. Combines reformatted output from Li & Kusche (2026) [5], with one variable per initialisation date (YYYY-MM) and six lead-time steps (months 0–5). The non-linear forecast component is computed as TWSC_full − TWSC_linear to match the detrended target used during training. References [1] Li et al.: Geophysical Research Letters, 48, e2021GL093492, https://doi.org/10.1029/2021GL093492, 2021. Data: https://datadryad.org/dataset/doi:10.5061/dryad.z612jm6bt [2] Hersbach et al.: ERA5 monthly averaged data on single levels, Copernicus CDS, https://doi.org/10.24381/cds.f17050d7, 2023. [3] Save, H.: CSR GRACE and GRACE-FO RL06 Mascon Solutions v02, https://doi.org/10.15781/cgq9-nh24, 2025. [4] Save et al.: Journal of Geophysical Research: Solid Earth, 121, 7547–7569, https://doi.org/10.1002/2016JB013007, 2016. [5] Li & Kusche: Water Resources Research, 62, e2025WR041710, https://doi.org/10.1029/2025WR041710, 2026. Data: https://doi.pangaea.de/10.1594/PANGAEA.973113