High-resolution soil temperature and soil moisture patterns in space, depth and time: An interpretable machine learning modelling approach

Soil temperature and soil moisture are key drivers of various soil ecological processes, which implies a significant importance of datasets including their variations in space, depth and time (4D). Current gridded products typically have a low resolution, either spatially or temporally. Here, we aim at modelling and explaining high-resolution soil temperature and soil moisture patterns in 4D for a 400 km${}^2$ study area in a heterogeneous landscape. Our target resolution of 10 m in space, 10 cm in depth, and 1 h in time allows capturing small-scale variations as well as short-term dynamics. We used multi-depth soil temperature and soil moisture measurements from 212 locations and linked them to 45 potential predictors, representing meteorological conditions, soil parameters, vegetation coverage, and landscape relief. We trained random forest models that were able to predict soil temperature with a mean absolute error of 0.93 °C and soil moisture with a mean absolute error of 4.64 % volumetric water content. Continuous model predictions enabled a comprehensive analysis of 4D patterns and confirmed that the selected resolution is essential to capture soil temperature and soil moisture variations at the landscape scale. In addition to a strongly pronounced annual cycle and recognisable influences on the diurnal cycle, there were significant differences between the land uses and patterns due to the relief, which also varied along the depth gradient. By applying interpretable machine learning techniques, we investigated the detailed influence of all drivers and discussed overlapping effects that led to the prediction patterns.