Bias-Corrected Gridded Soil Temperatures (North of 30°N) Between 1982-2023 at 0.05° Resolution

ECMWF Reanalysis 5th Generation Land (ERA5-Land) and Famine Early Warning Systems Network Land Data Assimilation System (FLDAS) both provide gridded soil temperatures globally at a resolution of ~9 kilometers (km). However, ERA5-Land soil temperatures exhibit warm biases over permafrost regions, and a median Root Mean Squared Error (RMSE) of between 1.6 Kelvin (K) and 2.3K, while FLDAS exhibits cold biases, and a median RMSE of between 2.5K and 4.5K. Thus in an uncorrected form, their soil temperatures are unsuitable for permafrost applications, or as boundary conditions for hydrological models. Here we investigated the use of a hierarchy of bias-correction techniques including mean bias subtraction (MBS), multiple linear regression (MLR), and random forest regression (RF) to bias-correct ERA5-Land and FLDAS soil temperatures. The MLR and RF models incorporated 10 predictors including soil depth, soil temperatures, 2 meter (m) air temperatures and snow water equivalent (SWE) from the reanalysis product, Moderate Resolution Imaging Spectroradiometer (MODIS) Normalized Difference Vegetation Index (NDVI), the month of the year, as well as geospatial information such as elevation, latitude and longitude. MLR and RF models were used to predict the observed soil temperature for each grid cell, and metrics were compared against in-situ soil temperature measurements from 2686 stations. It was found that RF greatly outperformed MBS and MLR, providing an average RMSE reduction of between 46 percent (%) to 77% relative to the uncorrected product soil temperatures.
This dataset accompanies Herrington, T., Erler, A. and Fletcher, C. (in Review), Theoretical and Applied Climatology. It includes two datasets of 0.05° bias-corrected gridded soil temperatures over the extratropical northern hemisphere for all land areas north of 30 degrees North (°N). The first dataset includes bias-corrected ERA5-Land soil temperatures for each of the Hydrology Tiled ECMWF Scheme for Surface Exchanges over Land (HTESSEL) model layers, between 1981-2023. The second dataset includes bias-corrected FLDAS soil temperatures for each of the Noah LSM model layers, between 1982-2023. The bias-correction utilizes random forest regression and 10 predictors including soil depth, the reanalysis soil temperature, sine and cosine transformations of the month of the year, the reanalysis air temperature, MODIS NDVI, reanalysis snow water equivalent (SWE), along with elevation, latitude and longitude.