Application of machine learning to improve reanalysis soil temperatures over the extratropical northern hemisphere

Reanalysis products provide spatially homogeneous coverage for a variety of climate variables. However, previous research has shown that soil temperature estimates in many reanalysis products have substantial biases; particularly in winter. Here we evaluate the performance of two products: ERA5-Land and FLDAS across the northern hemisphere and test a hierarchy of statistical techniques for bias correction. Both products provide high-resolution ( $$\simeq $$ 9km) estimates of vertically resolved soil temperature; however, ERA5-Land exhibits warm biases over permafrost regions, and a median RMSE of between 1.6K and 2.3K, while FLDAS exhibits cold biases, and a median RMSE of between 2.5K and 4.5K. Here we use multiple linear regression (MLR) and random forest regression (RF) for bias correction and compare them to mean bias subtraction (MBS) as a reference. The MLR and RF models employ 10 predictors, including soil depth, product soil temperature, air temperature, vegetation, snow cover, elevation, latitude and longitude. The RF model substantially outperforms MBS and MLR over all regions and latitudes, providing an average RMSE reduction (relative to the products’ soil temperatures) of 46% – 77% when the ground is snow-covered, and 56% – 64% during snow-free conditions. We introduce a bias-corrected soil temperature product, which provides gridded soil temperature data over the extratropical northern hemisphere between 1982 and 2023. This new data resource will be useful for a wide range of applications, including as an initialization condition for hydrological models, and as a tool to validate simulated soil temperatures.