Dataset associated with: Van Nieuwenhove H., Bechtold M., Lhermitte S., Desai A.R., De Lannoy G.J.M. "Can deep learning outperform mechanistic modeling of peatland water table dynamics?"

Dataset associated with: Van Nieuwenhove H., Bechtold M., Lhermitte S., Desai A.R., De Lannoy G.J.M. "Can deep learning outperform mechanistic modeling of peatland water table dynamics?" published in Machine Learning. Domain: Natural northern peatlands, 40°N–75°N, 170°W–100°ESpatial resolution: 9 km (EASEv2 M09 grid)Time period: 1990–2023, dailyVariable: Water level (WL) relative to ground surface (m); positive values indicate flooding. FILES----- lstm_ensemble_mean.nc Ensemble mean daily water level simulated by the LSTM model. The ensemble consists of six sub-models trained via blocked 6-fold cross-validation using in situ WL observations from 108 peatland monitoring wells (1999–2018). Inputs: rainfall (MERRA-2), reference evapotranspiration (ET0, Penman–Monteith), snow water equivalent (from PEATCLSM), and climatological LAI (MODIS/GEOLAND2). peatclsm.nc Daily water level (variable: zbar) simulated by PEATCLSM (NASA Catchment Land Surface Model with peatland modules), the process-based benchmark model used in the study. PEATCLSM parameters are literature-based and were not calibrated to the in situ data. NOTES------ Both files cover the same spatial domain and are provided for direct comparison.- Performance metrics reported in the paper are computed over non-frozen periods only.- For site-level validation statistics, see Table S2 of the supplementary material.