Skip to main content
ARS Home » Northeast Area » Beltsville, Maryland (BARC) » Beltsville Agricultural Research Center » Hydrology and Remote Sensing Laboratory » Research » Publications at this Location » Publication #335029

Research Project: Leveraging Remote Sensing, Land Surface Modeling and Ground-based Observations ... Variables within Heterogeneous Agricultural Landscapes

Location: Hydrology and Remote Sensing Laboratory

Title: Using machine learning to produce near surface soil moisture estimates from deeper in situ records at U.S. Climate Reference Network (USCRN) locations: Analysis and applications to AMSR-E satellite validation

item Coopersmith, E - University Of New Hampshire
item Cosh, Michael
item Bell, Jesse - National Oceanic & Atmospheric Administration (NOAA)
item Boyles, Ryan - North Carolina State University

Submitted to: Advances in Water Resources
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 12/1/2016
Publication Date: 12/15/2016
Citation: Coopersmith, E., Cosh, M.H., Bell, J., Boyles, R. 2016. Using machine learning to produce near surface soil moisture estimates from deeper in situ records at U.S. Climate Reference Network (USCRN) locations: Analysis and applications to AMSR-E satellite validation. Advances in Water Resources. 98:122-131.

Interpretive Summary: Many in situ networks monitor soil moisture at a variety of depths in the near surface soil column. These depths are not standardized between networks however. Satellite soil moisture observations require a near surface (5 cm) estimate for calibration and validation, but not all networks have a 5 cm estimate, such as the EcoNet in North Carolina. To meet this challenge, a model was applied to the EcoNet 10 cm soil moisture data series to produce a 5 cm estimate. This estimate was then compared to satellite data, providing a new resource for satellite validation in the eastern U.S. Validation was conducted using a select few in situ products which do have 5 and 10 cm sensors. This approach is useful for in situ network managers who need soil estimates, but are limited by sensor availability.

Technical Abstract: Surface soil moisture is critical parameter for understanding the energy flux at the land atmosphere boundary. Weather modeling, climate prediction, and remote sensing validation are some of the applications for surface soil moisture information. The most common in situ measurement for these purposes are sensors that are installed at depths of approximately 5cm. There are however sensor technologies and network designs which do not provide an estimate at this depth. If soil moisture estimates at deeper depths could be extrapolated to the near surface, in situ networks providing estimates at other depths would see their values enhanced. Soil moisture sensors from the Climate Reference Network (USCRN) were used to generate models of 5cm soil moisture, with 10cm soil moisture measurements and antecedent precipitation as inputs, via machine learning techniques. Validation was conducted with the available, in situ, 5cm resources. It was shown that a 5cm estimate, which was extrapolated, from a 10cm sensor and antecedent local precipitation, produced a root-mean-squared-error (RMSE) of 0.0215 m3/m3. Next, these machine-learning-generated 5cm estimates were also compared to AMSR-E estimates at these locations. These results were then compared with the performance of the actual in situ readings against the AMSR-E data. The machine learning estimates at 5cm produced an RMSE of approximately 0.03 m3/m3 when an optimized gain and offset were applied. This is necessary considering the performance of AMSR-E in locations characterized by high vegetation water contents, which are present across North Carolina. Lastly, the application of this extrapolation technique is applied to the ECONet in North Carolina, which provides a 10cm depth estimate as its most shallow soil moisture estimate. A raw RMSE of 0.028 m3/m3 was achieved, and with a linear gain and offset applied at each ECONet site, an RMSE of 0.013 m3/m3 was possible.