Surface soil moisture retrieval underpins climate modeling, agricultural management, and disaster prediction. Physical models, such as radiative transfer equations, offer strong interpretability but require numerous parameters that are difficult to calibrate, limiting their accuracy. Machine learning, in contrast, can model nonlinear relationships between soil moisture and environmental variables but depends on large, labeled datasets and lacks physical transparency. Because most soil moisture stations are concentrated in a few regions, it is challenging to perform accurate retrievals elsewhere. Due to these challenges, it is necessary to couple physical and data-driven approaches to achieve reliable, large-scale soil moisture estimation under sparse in situ data conditions.
Researchers from Wuhan University and the China University of Geosciences have developed a differentiable modeling framework that integrates physical and machine learning methods for large-scale soil moisture retrieval. The study (DOI: 10.34133/remotesensing.0367), published on October 1, 2025, in Journal of Remote Sensing, presents a novel model capable of generating 1-km resolution soil moisture maps across China using limited ground-based data. The new hybrid model enhances retrieval accuracy and generalization, offering an efficient solution for regions with few soil moisture monitoring sites or publicly available datasets. The team combined physical models—including the τ–ω radiative transfer model, Q–H model, Fresnel equation, and Mironov dielectric mixing model—with neural networks to form a fully differentiable retrieval framework. Unlike conventional physical models requiring empirical coefficients, this framework allows parameters to be optimized through gradient descent, bridging physical principles with data learning. The model was trained and validated using data from the Shandian River Basin and Naqu regions, representing contrasting climatic and topographic conditions. Results showed that the differentiable model achieved a correlation coefficient (R) of 0.925 and an unbiased RMSE of 0.035 m³·m⁻³ across all sites, rivaling state-of-the-art machine learning models while maintaining physical interpretability. It also outperformed traditional algorithms in spatial extrapolation, successfully capturing soil moisture variations in areas with distinct land cover and precipitation patterns. Moreover, the model’s retrievals closely matched SMAP satellite products, reproducing both spatial and temporal soil moisture dynamics across China.
“Our differentiable modeling approach represents a step toward uniting physical understanding with data-driven flexibility,” said Dr. Qiangqiang Yuan, corresponding author of the study. “It enables soil moisture retrieval with high accuracy even in data-sparse regions, without compromising physical realism. This framework not only advances remote sensing of soil moisture but also provides a blueprint for integrating physics and AI in Earth observation, allowing more robust environmental monitoring under diverse climatic and geographic conditions.”
The proposed differentiable framework demonstrates strong potential for improving the accuracy and consistency of soil moisture products worldwide, particularly in countries where in situ monitoring networks are sparse. Its high-resolution outputs can support precision agriculture, drought forecasting, and water resource management by providing timely and spatially detailed soil moisture information. Beyond soil moisture, the methodology can be extended to retrieve other surface parameters—such as evapotranspiration, vegetation water content, and land surface temperature—by coupling physical models with neural networks, paving the way for next-generation Earth observation systems that blend physics and artificial intelligence.
###
References
DOI
10.34133/remotesensing.0367
Original Source URL
https://doi.org/10.34133/remotesensing.0367
Funding information
This work was supported in part by the National Key Research and Development Program of China under grant 2022YFB3903403.
About Journal of Remote Sensing
The Journal of Remote Sensing, an online-only Open Access journal published in association with AIR-CAS, promotes the theory, science, and technology of remote sensing, as well as interdisciplinary research within earth and information science.