Machine Learning for Data Preparation: Improving Soil Moisture Filed

Supervised and unsupervised machine learning (ML) approaches have been shown to effectively detect patterns in large data sets. Given the large volume of available satellite data, it is a challenge to extract meaningful information in real time to improve forecasting model. We are constructing a test case to improve the soil moisture field with the data from the Advanced Baseline Imager (ABI) on the GOES-16.

Prior work has shown that soil state (soil moisture and temperature) has a notable impact on Quantitative Precipitation Forecast (QPF). It affects the amount of moisture in the boundary layer available for clouds and precipitation, and additionally plays a role in determining when and where convective initiation occurs. In the current Rapid Refresh (RAP) operational system, soil temperature adjustments were estimated experimentally, and are static per season. The RAP analysis is used to initialize the High Resolution Rapid Refresh (HRRR) for operational forecasts, particularly for better forecasts of mesoscale convective systems and QPF associated with them.

Through the development and implementation of a series of supervised and unsupervised ML algorithms, we expect to produce a more accurate, higher spatial resolution soil moisture product. The new field will be used in the data assimilation process for the Rapid Refresh (RAP) and the High Resolution Rapid Refresh (HRRR) weather forecast model. We believe that it will improve the initial estimation of soil state as part of the model data assimilation process. We expect this new soil state field to improve: QPF, Heat Flux, surface temperature, moisture and wind through the forecast period.