Soil moisture (SM) is vital for understanding the Earth’s water cycle, soil health, and climate change. Missions like SMAP and AirMOSS estimate SM, and more recently, GNSS-reflectometry systems like CYGNSS are used for this purpose. CYGNSS offers high temporal resolution and lower cost compared to traditional microwave observatories. Current SM estimation methods using CYGNSS fall into two categories: 1) Physics-Based (PB) models, which are accurate but depend on detailed geophysical parameters often unavailable globally; and 2) Black-box Machine Learning (ML) models, which are generally not constrained by physics-based models and can produce inconsistent results. We propose a novel physics-guided ML model with two modules: a PB forward model generating Delay Doppler Maps (DDMs) from SM, and an ML inverse model predicting SM from DDMs. The PB model selected is the Improved Geometric Optics with Topography (IGOT), and the ML model is a multi-layer perceptron, trained in two stages: pre-training and fine-tuning, both influenced by the PB forward model. Inspired by CycleGAN, the pre-training stage optimizes a cycle consistency loss function to minimize the differences between measured DDM and synthetic DDM generated by the PB forward model, and between arbitrary SM and predicted SM. This method addresses the in situ data scarcity challenge and encourages the ML model to act as an inverse of the PB forward model, producing consistent results. The fine-tuning stage minimizes the difference between predicted and in situ measured SM and incorporates the pre-training loss function as a regularizer. The ML models will benefit from using other satellite and ancillary data sources, including SMAP. Our hybrid approach leverages the ML model’s ability to deduce complex data relationships while ensuring physically consistent results. Additionally, it reduces the time complexity of SM estimation due to offline training and moderate complexity inference. Validation will be conducted using CYGNSS observations over three SoilSCAPE sites: San Luis Valley, CO; Jornada Experimental Range, NM; and Walnut Gulch Experimental Watershed, AZ. These sites offer diverse conditions, from smooth terrain and dry soil to complex topographies, monsoon seasons, and vegetation, providing a robust testing ground for the proposed method.
@article{Shokri_2024, title={Integrating Physics and Machine Learning for Soil Moisture Retrieval Using CYGNSS Observations}, url={http://dx.doi.org/10.22541/essoar.173532516.62762667/v1}, DOI={10.22541/essoar.173532516.62762667/v1}, publisher={Wiley}, author={Shokri, Parnia and Tsagkatakis, Grigorios and Melebari, Amer and Hodges, Erik and Moghaddam, Mahta}, year={2024}, month=dec }