Soil moisture monitoring networks can provide real-time and accurate soil moisture measurements; however, missing values and the lack of unified measurement depths across different networks impedes soil moisture applications at regional and national scales. Therefore, methods for vertical extrapolation of soil moisture, i.e., using shallow soil moisture measurements to estimate deeper soil moisture, are needed for standardizing measurements to a set of common depths. This study compared three methods, artificial neural network (ANN), linear regression (LR), and exponential filter (ExpF), for vertical extrapolation of soil moisture using data from the Oklahoma Mesonet. Based on our analysis of intra-annual variations in soil moisture, we divided each year into two seasons, warm and cool. Our results demonstrate that all methods perform better in the warm season than in the cool season, especially at deeper depths. The Kling–Gupta efficiency was used to assess the performance of each method. All methods had similar performance for near-surface extrapolation of soil moisture (top 25 cm). Although the accuracy of all models tended to decrease with depth, the ExpF outperformed the other methods at deeper depths. The soil water index (SWI) is preferred over volumetric water content as input to the ExpF. Incorporating air temperature and an antecedent precipitation index into the ANN and LR methods did not significantly improve their accuracy. We demonstrated that both ExpF and general LR can be used for SWI extrapolation at sites where only surface soil moisture data are available. Soil properties may be useful for further improving the accuracy of the general LR method.
ASJC Scopus subject areas
- Soil Science