This paper presents a machine learning (ML) framework to derive a quasi-global soil moisture (SM) product by direct use of the Cyclone Global Navigation Satellite System (CYGNSS)’s high spatio-temporal resolution observations over the tropics (within
latitudes) at L-band. The learning model is trained by using in-situ SM data from the International Soil Moisture Network (ISMN) sites and various space-borne ancillary data. The approach produces daily SM retrievals that are gridded to 3 km and 9 km within the CYGNSS spatial coverage. The performance of the model is independently evaluated at various temporal scales (daily, 3-day, weekly, and monthly) against Soil Moisture Active Passive (SMAP) mission’s enhanced SM products at a resolution of 9 km × 9 km. The mean unbiased root-mean-square difference (ubRMSD) between concurrent (same calendar day) CYGNSS and SMAP SM retrievals for about three years (from 2017 to 2019) is 0.044 cm
with a correlation coefficient of 0.66 over SMAP recommended grids. The performance gradually improves with temporal averaging and degrades over regions regularly flagged by SMAP such as dense forest, high topography, and coastlines. Furthermore, CYGNSS and SMAP retrievals are evaluated against 170 ISMN in-situ observations that result in mean unbiased root-mean-square errors (ubRMSE) of 0.055 cm
and 0.054 cm
, respectively, and a higher correlation coefficient with CYGNSS retrievals. It is important to note that the proposed approach is trained over limited in-situ observations and is independent of SMAP observations in its training. The retrieval performance indicates current applicability and future growth potential of GNSS-R-based, directly measured spaceborne SM products that can provide improved spatio-temporal resolution than currently available datasets.
This is an open access article distributed under the Creative Commons Attribution License
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.