Gated recurrent units for modelling time series of soil temperature and moisture: An assessment of performance and process reflectivity

Soil temperature and moisture are important variables controlling ecological processes, but continuous high-resolution data are rarely available. Therefore, we used the correlation with widely accessible meteorological variables, including air temperature and precipitation, to develop models that predict time series of soil temperature and moisture. To model high-resolution time series, predictor and target variables had a temporal resolution of 1 h. We tested the applicability of Gated Recurrent Units with time series from one exemplary site. The models showed a high predictability on the four years test set with a mean absolute error of 0.87°C for soil temperature and 3.20% volumetric water content for soil moisture. We further investigated the plausibility of the models by passing simplified synthetic data to the trained models and thereby proved their ability to reflect known processes. Finally, we showed the potential to apply the models to other sites and soil depths using transfer learning. • We trained Gated Recurrent Units to predict time series of soil temperature and soil moisture in an hourly resolution over 4 years. • The models could predict soil temperature with an MAE of 0.87°C and soil moisture with an MAE of 3.20% volumetric water content. • Model interpretation techniques showed that the models reflect known processes. • We showed the potential to apply the models to other sites and soil depths using transfer learning.