Extending the Soil Moisture Record of the Climate Reference Network with Machine Learning (Invited Presentation)

Soil moisture estimation is crucial for agricultural decision-support and a key component of hydrological and climatic research. Unfortunately, quality-controlled soil moisture time series data are uncommon before the most recent decade. However, time series data for precipitation are accessible at Climate Reference Network (CRN) locations since, in some cases, the 1940s. By adapting a simple bucket model that translates a precipitation time series into a soil moisture time series at the hourly scale and leveraging machine learning techniques to calibrate and correct its results, soil moisture calibrations are possible at any location with a sensed precipitation and soil moisture time series. Selecting a model that does not require initial conditions to deliver predictions facilitates the extension of these soil moisture time series to historical periods at the same locations for which only precipitation data are available. The success of this approach can be verified by calibrating these models using a subset of available years and validating the results using previous years at which both precipitation and soil moisture data are available. This procedure is possible at SCAN sites at which soil data records are often available beginning in 2002. Demonstrating the feasibility of backwards forecasting suggests that at many CRN sites for which extensive historical precipitation records are available, soil moisture records can be extended to match.