Soil moisture inversion in arid areas by using machine learning and fully polarimetric SAR imagery

Soil moisture is one of the most important variables to affect the water cycle and energy balance during the evolution of regional ecosystem in arid areas. However, accurate monitoring of soil moisture is still a challenging task, due to the spatial and temporal heterogeneity. Radar remote sensing has widely been expected to be one of the most effective technologies in regional soil moisture monitoring. Fully polarimetric SAR (PolSAR) can also provide abundant polarized information for different machine learning algorithms to retrieve soil moisture in various regions. However, such research is still lacking in most arid areas, together with the specific evaluation on the performance of different machine learning algorithms. This study aims to retrieve the soil moisture in arid areas using the PolSAR parameters and various machine learning algorithms. The study area was selected as Juyanze region located in the southeast of Ejina banner of Inner Mongolia in western China. Based on Radarsat-2 imagery, radar variables were set as the extracted backscattering coefficients (BC) using the standard intensity and phase processing, while the multiple polarimetric parameters that derived from Cloude-Pottier decomposition (CPD) and Yamaguchi decomposition (YD). The parameter correlation and importance were also analyzed after that. Then, 21 soil moisture inversion models were established using three machine learning algorithms, namely Random Forest (RF), Support Vector Machine (SVM), and Back Propagation Artificial Neural Network (BP-ANN). Model performance was further evaluated using 10-fold cross-validation. Finally, the optimal model was achieved to inverse soil moisture in the study area, where the spatial distribution pattern was analyzed. The results show that: 1) The average scattering angle presented the most prominent influence on the inversion accuracy, followed by entropy and anisotropy among all the variables. Moreover, cross-polarized backscattering coefficients made much more contribution to the model accuracy, compared with the co-polarized backscattering coefficients. The importance of even scattering and volume scattering was remarkably higher than that of surface scattering and helix scattering. Parameters derived from CPD made outstanding contributions to the retrieval, where the importance scores and correlation coefficients were much higher than those of backscattering coefficients and parameters derived from YD. 2) The developed models of soil moisture inversion under the combined scheme of various variable types performed better than those built solely on single variable type in all three machine learning algorithms, indicating that the combined scheme greatly improved the accuracy of models.3) RF model was more suitable for soil moisture inversion in arid areas,compared with SVM and BP-ANN, according to the determination coefficient R2 and the root mean square error (RMSE). The model performed best using BC + CPD scheme as input variables. The validation set R2 and RMSE were 0.78 and 6.60%, respectively, with the standard deviation of R2 and RMSE of 0.15 and 1.95%, respectively. Consequently, 89% moisture variation can be explained by this optimal model. 4) Generally speaking, soil moisture in the study area maintained at a low level, and the average soil moisture content was 8.83%. Moisture content around the Swan Lake and the center of Paleolake was obviously higher than other areas. The inversion data conformed greatly to the actual situation, indicating a great potential to soil moisture inversion in arid areas.