Random forest to predict and forecast subsurface soil moisture in agricultural fields
Abstract
The importance of soil moisture is recognized globally because it controls hydrological processes that are relevant to agriculture and climate studies. Currently, estimation of subsurface soil moisture is largely accomplished using physical models, which are based on flow and transport equations across the vadose zone. However, with the complexity of the processes operating in the vadose zone as well as their interactions with each other, parameterizing all the relevant processes is quite a challenge. This complexity is further enhanced by the spatio-temporal changes in soil and vegetation properties which demand model parameters to be dynamic. Alternatively, purely data-based methods for subsurface soil moisture estimation is still limited. This is despite the availability of datasets from numerous networks established since the last decade. These datasets serve, among others, for calibrating and validating physical models and retrieval methods from satellites. They have also been utilized to understand the spatio-temporal soil moisture dynamics. In this study, we explore the utility of a machine learning method with Random Forest (RF) as an approach for predicting and forecasting daily subsurface soil moisture using time series datasets from selected stations in the Twente and Raam soil moisture networks in the Netherlands. We divided the total dataset into 70% training and 30% validation data. For prediction, the training data were randomly selected and the validation data was the remainder from the total dataset. For forecasting, the training data consists of the first 70% of the time series while the validation data is the last 30% of the time series. Daily meteorological data, surface soil moisture, and MODIS-derived Leaf-Area Index served as predictors in the RF algorithm. This set of predictors are also regularly used for simulating subsurface soil moisture using physical models. The calculated accuracy metrics for the prediction of subsurface soil moisture yielded an RMSE: 0.04 - 0.08 m 3 m -3 , bias: 0.002 - 0.07 m 3 m -3 , unbiased RMSE: 0.03 - 0.08 m 3 m -3 , and Spearman's correlation: 0.86-0.98. While prediction showed promising results, forecasting of subsurface soil moisture resulted lower accuracies: RMSE: 0.05 - 0.23 m 3 m -3 , bias: -0.16 - 0.01 m 3 m -3 , unbiased RMSE: 0.05 - 0.17 m 3 m -3 , and Spearman's correlation: 0.61-0.97. The RF model performed poorly for forecasting of extreme soil moisture conditions (e.g very wet and very dry), especially if these conditions are not captured in the training data. Overall, results show that a purely data-based method can reasonably predict and forecast subsurface soil moisture. This can be an alternative approach in cases where subsurface soil moisture the parameters values for processes needed in a physical models are absent or are very hard to obtain.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2019
- Bibcode:
- 2019AGUFM.H31I1830C
- Keywords:
-
- 1805 Computational hydrology;
- HYDROLOGY;
- 1875 Vadose zone;
- HYDROLOGY;
- 1879 Watershed;
- HYDROLOGY;
- 1942 Machine learning;
- INFORMATICS