Continental-scale prediction of soil organic carbon using optical-thermal-microwave satellite observation synergies

The advancements in cloud computing platforms (such as the Google Earth Engine (GEE)) and the optical, radar, and thermal infrared (TIR) observations hosted on these platforms significantly enhance satellite-based digital mapping of soil organic carbon (SOC). However, the literature generally lacks consensus on the effects of different optical, TIR, and radar sensors, as well as radar sensor configurations, on the digital mapping of SOC. In this study, we aimed to investigate the effects of these three satellite sensors and radar sensor configurations on the digital mapping of SOC at two soil depths (0–20 and 20–100 cm) on the continental (Asian) scale using long-term satellite observation data from the GEE. Models to predict SOC at two depths were developed by constructing eleven prediction scenarios with different data configurations using long-term optical (MODIS), thermal (Landsat-8), and radar (C-band Sentinel-1 and L-band PALSAR-1/2) observations combined with the World Soil Information Service (WoSIS) soil profile database and two machine learning methods, with performance evaluated through cross-validation. The results indicated that the SOC predictions at both depths were influenced by the machine learning algorithm, sensor type, and radar sensor configuration. The model trained using single-polarization data yielded the poorest performance, whereas cross-polarization outperformed copolarization for radar data under both bands; incorporating backscatter information obtained from more polarization modes and operating frequencies into the model improved the predictions. Specifically, dual‐band integration increased R² from 0.41 to 0.46 (0–20 cm) and 0.36–0.40 (20–100 cm) compared to L -band data alone. When training models independently using optical, TIR, and radar data, the radar-based model performed the worst, whereas the models based on TIR and optical data exhibited the best performance in terms of SOC prediction at depths of 0–20 cm (R² = 0.56) and 20–100 cm (R² = 0.46), respectively. Compared with the models developed with only one or two imaging techniques, the models developed with all three imaging techniques performed best across all the scenarios, obviously improving the predictions, and explained approximately 62 % of the SOC variability at the 0–20 cm depth and 53 % at the 20–100 cm depth. Importance analysis indicated that long-term optical, TIR, and radar observations were important in our models at different depths. The maps predicted by the three imaging techniques, which varied between depths and exhibited high spatial heterogeneity, were broadly similar in large-scale spatial gradients but differed mainly in fine-scale spatial detail and predicted magnitude. • Long-term optical, radar, and thermal imagery were important for models at different depths. • The modeling output of SOC was affected by the prediction method, sensor type, and sensor parameters. • Incorporating backscatter information obtained at more polarizations and frequencies improved the predictions. • The model developed using the three imaging techniques performed best. • The GEE platform can effectively integrate optical-radar-thermal data to support continental-scale SOC mapping.