*3.3. GNSS-R Soil Moisture Retrieval*

To analyze the influence of incoherent observations on the GNSS-R land surface SM inversion in the previous SM retrieval method, we compared two retrieval configurations: using 19 months CYGNSS land observations and screened coherent data for retrieval model evaluation with k-fold cross-validation approach, where k = 5. Since the global SM value in most areas of the land is generally small in a year, the PDF of the monthly SMAP SM data in 2018 is presented in Figure 7a, and the maximum probability density of SM is 0.06 cm3/cm3. To further evaluate the performance of the established SM model over the high-humidity areas, the accuracy of the inversion model is evaluated when the referenced SM value is greater than 0.1 cm3/cm3.

**Figure 7.** Monthly SMAP soil moisture probability density function (**a**) and density scatterplot of GNSS-R derived soil moisture and surface SM reference values in a split of k-fold cross-validation (**b**).

Using the SM inversion method introduced in Section 2.4, Table 2 summarizes the performance of the two models established from two training datasets. When all CYGNSS land observations are used for modeling, the cross-validation model bias, mean absolute error (MAE) and root-mean-square error (RMSE) are −0.0003 cm3/cm3, 0.0274 cm3/cm3, and 0.0416 cm3/cm3, respectively. The inversion results with the distinguished coherent observation training dataset constructed retrieval model show that the bias, MAE, and RMSE are −0.0003 cm3/cm3, 0.0269 cm3/cm3, and 0.0408 cm3/cm3, respectively. The model performance between the two strategies is very close. When the SM reference values are greater than 0.1 cm3/cm3, the model accuracy of the two methods is 0.0569 cm3/cm3 and 0.0564 cm3/cm3, respectively, and the inversion results did not show a big difference. Figure 7b shows the density scatterplot between SMAP reference SM and GNSS-R-derived SM generated from a split of k-fold cross-validation with the coherent observation established model. The red line represents the linear fitting line; the predicted SM shows an overall fairly good agreement with the SMAP SM, all CYGNSS land data retrieved SM show

an identical situation. Figure 8 presents the coherent inversion accuracy at each grid pixel with k-fold cross-validation. The analysis shows that CYGNSS incoherent observations will not cause any noticeable SM spatial inversion accuracy differences compared to the coherent results, so it is not given here.


**Table 2.** Soil moisture retrieval model evaluating with k-fold cross-validation (unit: cm3/cm3).

**Figure 8.** Root-mean-square error (RMSE) of CYGNSS retrieved land surface SM at each grid.
