Quasi-global machine learning-based soil moisture estimates at high spatio-temporal scales using CYGNSS and SMAP observations

Authors: LEI, F. - Mississippi State University; SENYUREK, V. - Mississippi State University; KURUM, M - Mississippi State University; GURBUZ, A.C. - Mississippi State University; BOYD, D. - Mississippi State University; MOORHEAD, R. - Mississippi State University; Crow, Wade; EROGLU, O. - National Center For Atmospheric Research (NCAR)

Submitted to: Remote Sensing of Environment
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 4/4/2022
Publication Date: 4/20/2022
Citation: Lei, F., Senyurek, V., Kurum, M., Gurbuz, A., Boyd, D., Moorhead, R., Crow, W.T., Eroglu, O. 2022. Quasi-global machine learning-based soil moisture estimates at high spatio-temporal scales using CYGNSS and SMAP observations. Remote Sensing of Environment. 276:113041. https://doi.org/10.1016/j.rse.2022.113041.
DOI: https://doi.org/10.1016/j.rse.2022.113041

Interpretive Summary: Accurate monitoring of soil moisture availability is important for a range of agricultural applications including drought monitoring, irrigation scheduling, and numerical weather prediction. Recent advances in remote sensing have enabled new possibilities for providing such monitoring using spaceborne sensors. However, existing remote sensing products are still hampered by a range of limitations including poor spatial resolution and retrieval noise. This paper describes the novel use of a machine learning technique to improve the temporal coverage and spatial resolution of satellite-based soil moisture estimates. By providing more-frequent and higher-resolution estimates of soil moisture availability in agricultural regions, this approach will eventually be used to improve our ability to track agricultural drought and forecast irrigation demand.

Technical Abstract: Global soil moisture mapping at high spatial and temporal resolution is important for various meteorological, hydrological, and agricultural applications. Recent research shows that the land surface reflected forward scattering Global Navigation Satellite System (GNSS) signals at L-band can convey high-resolution land surface information, including surface soil moisture. However, these signals are often affected by complex land surface characteristics and the bistatic nature of the GNSS-Reflectometry (GNSS-R) technique, resulting in a nonlinear relationship between the signals and surface soil moisture. In this work, a machine learning (ML) approach is used to map quasi-global soil moisture using bistatic reflectance observations acquired from the recently launched Cyclone GNSS (CYGNSS) mission. Specifically, several land surface parameters are obtained from remote sensing products and integrated with SMAP enhanced soil moisture retrievals to facilitate the daily quasi-global CYGNSS soil moisture mapping at 9 km. The ML algorithm is shown to be suitable for retrieving soil moisture from CYGNSS with the year-based cross-validation unbiased root-mean-square-difference of 0.0395 cm3/cm3 and 0.0320 cm3/cm3 for all quasi-global grids and regions with vegetation water content less than 5 kg/m2, respectively. Based on an independent evaluation against in situ measurements, CYGNSS has shown a comparable performance with respect to SMAP over more than 100 ground validation sites. CYGNSS also shows a similar spatial variability with SMAP across different seasons. Moreover, through a robust triple collocation technique, the accuracy of CYGNSS is demonstrated with high correlations over moderately vegetated regions. Therefore, the derived CYGNSS soil moisture estimates can supplement the current global soil moisture database and provide more frequent retrievals at 9 km.