An RTE-guided mixture-of-experts framework for AMSR2 passive microwave soil moisture retrieval

Systematic biases remain evident in the JAXA AMSR2 soil moisture product, particularly in densely vegetated and arid regions where vegetation, surface roughness, and soil emission interact nonlinearly. This study developed an RTE-guided mixture-of-experts (MoE) framework for AMSR2 soil moisture retrieval by combining iterative proxy-label refinement with a gated end-to-end MoE network trained on 14-channel brightness temperatures. Global data from 2020 were used for proxy-label refinement and model training. Temporally independent AMSR2 data from 2024 were used for cross-year proxy-reference consistency evaluation, and representative seasonal datasets from 2023 were used to assess seasonal robustness. Independent validation was conducted using International Soil Moisture Network (ISMN) in situ observations. In the pooled ISMN comparison, the proposed method reduced MAE from 0.048 to 0.034 m 3 /m 3 and RMSE from 0.068 to 0.046 m 3 /m 3 relative to the JAXA product, with ubRMSE = 0.046 m 3 /m 3 , R = 0.825, and R-anom = 0.709. The refined-label MoE also outperformed the MoE trained directly on JAXA-initialized labels, reducing MAE from 0.039 to 0.034 m 3 /m 3 and RMSE from 0.055 to 0.046 m 3 /m 3 . Regionally, MAE and RMSE decreased from 0.058 to 0.041 m 3 /m 3 and from 0.093 to 0.055 m 3 /m 3 over the tropical LABFLUX network, and from 0.029 to 0.025 m 3 /m 3 and from 0.047 to 0.033 m 3 /m 3 over the semi-arid REMEDHUS network. In the 2024 proxy-reference consistency evaluation, the 14-channel configuration achieved MAE = 0.012 m 3 /m 3 and R = 0.936, indicating high emulation fidelity rather than independent accuracy. These results suggest that the proposed framework improves agreement with in situ observations while maintaining strong proxy-reference consistency.