Field-scale soil moisture dynamics predicted by deep learning

Soil moisture plays a critical role in land–atmosphere interactions. Prediction of its dynamics is still a grand challenge. While in-situ measurements using sensors offer highly temporally resolved and accurate information compared to satellite observations, existing sensor networks are sparse and scarce. Here we propose a deep learning model for bridging the gap between infrequent satellite observations and sparse in-situ sensor network to improve near-term soil moisture predictions. The Long Short-Term Memory (LSTM)-based deep learning model was used to forecast soil moisture dynamics using soil parameters and climatic variables (e.g. air temperature, relative humidity, pressure, wind speed, turbulent fluxes, solar and terrestrial waves) collected from a dense network of sensors in a field located in Germany in an area of about 20 hectares. The dynamic time-lagged cross-correlation between soil moisture and other co-located soil and climatic features was calculated and a set of optimal predictors for training the LSTM model was selected. To efficiently learn the long-term dependency of soil moisture on its historical trends and to improve the prediction capability of the model, we optimized the LSTM structure, hyperparameters, and the size of the sliding window based on the goodness of fit ($R^2$ score) of the model. We also examined the feasibility of employing the model developed using temporal data from one location for the prediction of soil moisture at other locations across the landscape. The results illustrate the robustness and efficiency of the proposed model for the spatio-temporal prediction of soil moisture. Plain Language Summary Understanding and monitoring soil moisture dynamics is crucial affecting ecosystem health, climate and extreme weather patterns, and the agricultural sector. However, predicting the temporal and spatial variation of soil moisture is challenging because of the complex interactions between the land and atmosphere. While soil moisture measurement with in-situ ground-based sensors provide a high level of temporal frequency in comparison to satellite data, the implementation of dense monitoring networks to capture spatial variability of soil moisture is not economically viable. To address this problem, we utilized machine learning techniques to predict temporal and spatial variation of soil moisture using data we measured in a field in Germany. The developed model was examined against the experimental data with the results illustrating that AI-based solutions could offer a powerful tool to predict soil moisture dynamics.