Yan et al. (2025) A Modified Hierarchical Vision Transformer for Soil Moisture Retrieval From CYGNSS Data

⚠️ Warning: This summary was generated from the abstract only, as the full text was not available.

Identification

• Journal: Water Resources Research
• Year: 2025
• Date: 2025-12-27
• Authors: Qingyun Yan, Yuhan Chen, Yuanjin Pan, Shuanggen Jin, Weimin Huang
• DOI: 10.1029/2024wr039476

Research Groups

[Information not available in the abstract.]

Short Summary

This research introduces a novel deep learning framework, multi‐head self‐attention‐aided vision Transformer (MSA‐ViT), for soil moisture retrieval using Cyclone Global Navigation Satellite System (CYGNSS) data. The MSA-ViT model integrates physical understanding with deep learning to capture nonlinear interactions, demonstrating superior performance over conventional and established deep learning models, and improving upon existing CYGNSS L3 products.

Objective

• To introduce and validate a new deep learning framework, multi‐head self‐attention‐aided vision Transformer (MSA‐ViT), for soil moisture (SM) retrieval using Cyclone Global Navigation Satellite System (CYGNSS) data.
• To assess the sensitivity of CYGNSS reflectivity to SM, establishing a physical linkage through coherent scattering theory.
• To integrate physical understanding with deep learning to capture nonlinear interactions between SM, surface roughness, and vegetation attenuation.
• To demonstrate the MSA-ViT model's superior performance compared to conventional and established deep learning models, and its improvement over the current CYGNSS L3 product.
• To validate the model's robustness and applicability across large watersheds and diverse spatiotemporal scales for different ecosystem types.

Study Configuration

• Spatial Scale: Large watersheds, diverse ecosystem types.
• Temporal Scale: Data from January 2020 to December 2024 (5-year period); observations aggregated over multiple temporal scales (3–60 days); initial training with 10-day averaged data; evaluation with 3-day retrievals.

Methodology and Data

• Models used:
  • Multi-head self-attention-aided vision Transformer (MSA-ViT) (proposed)
  • Linear regression (comparative)
  • Shallow neural networks (comparative)
  • Other established deep learning models (comparative)
• Data sources:
  • Cyclone Global Navigation Satellite System (CYGNSS) reflectivity data
  • Soil Moisture Active Passive (SMAP) data
  • CYGNSS L3 SM V3.2 product (for comparison)
  • International Soil Moisture Network (ISMN) measurements (in situ)
  • Global Precipitation Measurement (GPM) records

Main Results

• A strong physical linkage between CYGNSS reflectivity and soil moisture was demonstrated through coherent scattering theory.
• The proposed MSA‐ViT model effectively integrates physical understanding with deep learning to capture nonlinear interactions among soil moisture, surface roughness, and vegetation attenuation.
• The MSA‐ViT‐based approach significantly outperformed conventional techniques (linear regression, shallow neural networks) and other established deep learning models in soil moisture retrieval.
• The model's 3-day retrievals showed consistent alignment with Soil Moisture Active Passive (SMAP) soil moisture data and accurately reflected seasonal variability.
• The MSA-ViT model demonstrated improved precision and coverage compared to the current CYGNSS L3 product.
• Comprehensive validation across large watersheds and diverse spatiotemporal scales confirmed the model's robustness and applicability across various ecosystem types.

Contributions

• Introduction of a novel deep learning framework (MSA-ViT) specifically designed for soil moisture retrieval using CYGNSS data, which uniquely integrates physical understanding with deep learning.
• Demonstration of superior performance of the MSA-ViT model over existing conventional and deep learning methods for soil moisture retrieval.
• Significant improvement in both precision and coverage for soil moisture products compared to the current CYGNSS L3 product.
• Comprehensive validation across diverse spatiotemporal scales and ecosystem types, enhancing the utility and applicability of CYGNSS data for global soil moisture monitoring.

Funding

[Information not available in the abstract.]

Citation

@article{Yan2025Modified,
  author = {Yan, Qingyun and Chen, Yuhan and Pan, Yuanjin and Jin, Shuanggen and Huang, Weimin},
  title = {A Modified Hierarchical Vision Transformer for Soil Moisture Retrieval From CYGNSS Data},
  journal = {Water Resources Research},
  year = {2025},
  doi = {10.1029/2024wr039476},
  url = {https://doi.org/10.1029/2024wr039476}
}