**4. Discussion**

GNSS-R coherent and incoherent observations have different sensitivities to the land SM values [9]. According to the classification results with the defined estimator in this study, 6.2% of the measurements in the CYGNSS land observations have a high possibility controlled by the incoherent scattering field. In addition, the PDFs of reflectivity calculated from the land surface coherent and incoherent observations do show distribution differences, as shown in Figure 9. It should be noted that if the DDM scattering power is dominated by incoherent components, NBRCS is commonly picked as the fundamental quantity, which is calculated according to [21]. Since most of the previous studies ignored incoherent scattering, namely the counterpart reflectivity is directly calculated by Equation (8), it is reasonable to use this equation to calculate incoherent reflectivity and analyze their influence on CYGNSS SM retrieval in this paper. The experiments show that extra incoherent observations have no obvious effect on the final CYGNSS SM retrieval with space-time averaging combined with the linear regression method. To further validate this conclusion, the threshold of NIDW-derived TES is set to −0.5 to improve the confidence of discriminated coherent DDM, where the probability of false alarm is only 0.01. It also can be considered that the contribution of the incoherent component is very small in screened coherent observations. At this point, the coherent DDM accounts for 75.8% of CYGNSS land measurements. The bias, MAE, and RMSE of final inversed soil moisture are −0.0003 cm3/cm3, 0.0265 cm3/cm3, and 0.0403 cm3/cm3, respectively. The RMSE is reduced by 3.1% compared to the constructed model with assuming all coherent land observations. When the reference SM value is greater than 0.1 cm3/cm3, the inversion bias is −0.0145 cm3/cm3, MAE values is 0.0416 cm3/cm3, and RMSE is 0.0558 cm3/cm3.

**Figure 9.** The probability density function of coherent and incoherent reflectivity derived from CYGNSS land observations in January 2018.

The inversion accuracy of the aforementioned GNSS-R space-time averaging SM retrieval methods with two different training datasets is similar because the magnitude and number of incoherent reflectivity are smaller when compared to coherent reflectivity, and the spatial average processing will further mitigate its influence. However, there is no doubt that coherence classification methods play a key role in future GNSS-R land detection. The inversion model can be directly established at the individual specular point with improved high-quality and high spatial resolution observations, which provides in the following dedicated spaceborne GNSS-R land remote sensing missions, and also contributes to other land applications, such as inland water system detection, biomass detection, and wetland extent determination. Another noteworthy issue is that the established GNSS-R SM inversion model tends to underestimate the surface soil moisture when the land SM over 0.3 cm3/cm3, while most of the previous studies also show the same problem. Since most training samples are concentrated in the lower SM range, the regression model is more affected by this part of the data. Therefore, there should be a better weighting strategy to solve this problem in future work.
