Objectives Global navigation satellite system-reflectometry (GNSS-R ) is a new method for measuring soil moisture with rich signal sources, which utilizes GNSS reflected signals for detecting the physical parameters of the reflector with cheap cost and high spatial resolution.

Methods Considering the fact that in the traditional GNSS-R soil moisture retrieval, the data of land cover types is ignored or simplified, this paper makes detailed labels about the land cover types based on the definition of international geosphere biosphere programme (IGBP) considering the similarities and differences between different land cover types, and uses back propagation (BP) neural network to establish soil moisture inversion models. The main work is to replace the IGBP number labels with refined feature labels. After the extraction, matching and quality control of cyclone global navigation satellite system (CYGNSS) and soil moisture active passive (SMAP) data, the dataset is randomly divided into training set and testing set. Then BP algorithm is used for training neural networks to establish soil moisture retrieval model.

Results The correlation coefficient between retrieved soil moisture and reference value is 0.85, and the root mean square error (RMSE) is 0.060 cm³/cm³. Meanwhile, the soil moisture obtained by CYGNSS and SMAP have good spatial and temporal consistency, but the inversion performance varies on the different land cover types. On the barren or sparsely vegetated land type, the RMSE is smallest and on the evergreen ncedlcleaf forest land type, the RMSE is largest. Meanwhile on the shrublands, grasslands and croplands land types, the inversion effect is better than permanent wetlands, urban/built-up and cropland/natural vegetation mosaic land types. Compared with the existing IGBP numbering and labeling method, the accuracy has been improved by 7.7%. The precision improvement of deciduous broadleaf forest, closed shrublands and woody savannas types is significant.

Conclusions The proposed method of refined land type digitization can effectively retrieve soil moisture without relying on other auxiliary data such as vegetation opacity and surface roughness, and it is feasible and effective.