When Statistical Learning Violates Physics: An Extensive Mismatch in Soil Temperature–Depth Relationships in Global Maps of Soil Temperature

Global gridded soil temperature datasets are important to understand and explain spatial patterns and processes in many life and environmental sciences, but products based on in situ measurements are still available to a very limited extent. Global maps of soil temperature at a 1-km² resolution for two depth levels of 0-5 cm and 5-15 cm were therefore an important step in bridging this gap. However, there are on average 26% of suspicious grid cells, and 7%-46% for individual considered soil bioclimatic variables, that show reversed patterns between the two depth levels in terms of soil temperature physics, with more pronounced temperature amplitudes, minima, and maxima at the deeper level, which has no reasonable physical explanation. This mismatch is most probably due to the fact that soil temperature grids for the two depth levels were generated independently using machine-learning models based on distinctive and spatially averaged sets of in situ soil temperature measurements for differing time periods. While the application potential of the maps remains enormous, and they can still be used for most soil-related applications, it can definitely be recommended that the two depth levels be used separately. The study also suggests that data consistency should be prioritized over maximizing the volume of data used when producing soil temperature grids at multiple depth levels using statistical learning methods based on in situ measurements.