Greenland ice sheet ice and firn temperature reconstruction at 10 m depth, 1950-2022

This dataset contains monthly grids of 10 m subsurface temperatures (T10m) for the Greenland ice sheet between 1954 and 2022. These grids have been produced by an Artificial Neural Network (ANN) which take as input ERA5 snowfall and air temperatures and is trained on more than 4500 observations of T10m from multiple sources. For a full description of the observation dataset, ANN training and performance, please see: Vandecrux, B., Fausto, R. S., Box, J., Covi, F., Hock, R., Rennermalm, A., Heilig, A., Abermann, J., van As, D., Bjerre, E., Fettweis, X., Smeets, P.C.J.P., Kuipers Munneke, P., van den Broeke, M., Brils, M., Langen, P.L., Mottram, R., Ahlstrøm, A.: Historical snow and ice temperature compilation documents the recent warming of the Greenland ice sheet, manuscript in development, 2023 The dataset is composed of two netcdf files: - T10m_prediction.nc which contains the ANN's prediction - T10m_uncertainty.nc which contains the ANN's estimated uncertainty Both files have latitude and longitude (WGS84) as coordinate reference system (CRS), a monthly temporal resolution and a spatial resolution of 0.1°x0.1°. The uncertainty is calculated from spatial cross-validation: We devide the ice sheet into 10 geographic regions which each contain between 95 and 1280 observations, meaning from 2% to 28% of the observation dataset. We train 10 ANN models, each of them ignoring one of these regions and therefore not learning from the observations therein. For a given location and month, the standard deviation between these 10 cross-validation models is taken as the ANN uncertainty. The best model, which uses all available observations, is used to produce the T10m_prediction.nc.