Airborne L-band SAR-assisted co-kriging for sequential multi-depth soil moisture retrieval

Soil moisture retrieval below the immediate surface remains difficult because vegetation attenuation, limited microwave penetration, and point–pixel scale mismatch constrain cultivated-layer estimates. This study developed a sequential multi-depth soil moisture inversion framework for a 7.0 km × 3.5 km agricultural area in the Huaihe River Basin, China, by integrating Sentinel-1/2 imagery, airborne L -band synthetic aperture radar ( L -SAR), topographic descriptors, and measurements from 106 sites. Soil moisture at 0–8 cm was retrieved using a machine-learning model after fold-local screening based on Pearson correlation, mutual information, and standardized Elastic-Net coefficients. Soil moisture at 8–17 and 17–25 cm was then estimated using an auxiliary-conditioned co-kriging scheme driven by preceding-layer out-of-fold (OOF) predictions and L -SAR-derived auxiliary information. Ridge achieved the best performance with a 25-feature predictor set (median R = 0.79; RMSE = 2.63%). Joint co-kriging produced R = 0.68 and RMSE = 3.36% at 8–17 cm and R = 0.59 and RMSE = 3.88% at 17–25 cm. Spatial validation showed stable 0–8 cm performance but lower and more separation-sensitive performance at deeper layers, especially 17–25 cm under minimum-separation Block-CV. Maps showed coherent depth ordering and local variation but only partial agreement with coarse soil-property products, which could not independently validate fine-scale moisture patterns. However, L -SAR was more informative within 0–17 cm but added little at 17–25 cm, especially under vegetation, so it should be treated as a spatial auxiliary rather than a direct sensor of deep soil moisture. Together, these results support the proposed framework under the tested conditions and provide a basis for future irrigation decision-making.