An Oasis Soil Moisture Inversion Model Using ALOS-2 and Landsat 8 Data

Objectives Integration of machine learning and multi-source data becomes a hot topic in soil moisture inversion, where relatively few studies are performed on L-band synthetic aperture radar （SAR） imagery.

Methods ALOS-2 PALSAR-2 and Landsat 8 images of Ejina Oasis are used to extract the radar and optical characteristic parameters which are then screened according to the importance score. Random forest is adopted to establish different soil moisture inversion models based on radar, optical, and radar-optical integrated parameters. Model accuracies are evaluated and soil moisture content in Ejina Oasis is inversed.

Results Compared with C-band, L-band SAR data is more sensitive to soil moisture content in arid desert oasis. With regard to radar characteristic parameters, surface and volume scattering components have higher important scores, while dihedral and helix scattering component are less important. As for optical characteristic parameters, vegetation water supply index takes the most important place while the enhanced vegetation index is the least important one. The determination coefficient R² and root mean square error (RMSE) of radar characteristic parameter scheme are 0.67 and 2.16%, respectively. The accuracy of optical characteristic parameter scheme model is generally low and the accuracy is equivalent, with R² and RMSE about 0.5 and 2.47%, respectively. R² and RMSE of the optimal radar-optical integrated parameter inversion model are 0.72 and 1.99%, respectively. Compared with either single data source, R² is increased by 7.46% and 38.4%, while RMSE is decreased by 8.54% and 22.6%.

Conclusions The research proves that the random forest model based on multi-source data fusion has higher prediction accuracy and better applicability in arid desert oasis area.