Field-scale soil moisture estimated from Sentinel-1 SAR data using a knowledge-guided deep learning approach

Technical Summary: Field-Scale Soil Moisture Estimation from Sentinel-1 SAR Data Using a Knowledge-Guided Deep Learning Approach

Problem Statement
Accurate, high-resolution estimation of field-scale soil moisture (SM) is critical for agricultural and environmental studies but remains challenging. While passive microwave sensors offer reliable, all-weather SM retrieval, their coarse spatial resolution (tens of kilometers) limits applicability at the field scale. Conversely, C-band Synthetic Aperture Radar (SAR) sensors, such as Sentinel-1, provide high spatial resolution (10–30 m) but face inherent complexities in signal interpretation due to the multiplicative interactions between radar backscatter and surface characteristics like vegetation cover, surface roughness, and soil texture.

Existing retrieval methods generally fall into two categories with distinct limitations:

1. Empirical/Semi-physical Models (e.g., Water Cloud Model - WCM): These decompose total backscatter into soil and vegetation contributions based on theoretical relationships. However, they often rely on simplified assumptions (e.g., linear relationships between SM and soil backscatter) and site-specific parameterization, which restricts their generalizability across diverse agricultural landscapes.
2. Purely Data-Driven Models (e.g., Deep Neural Networks): While capable of capturing complex non-linear relationships, these models are frequently criticized for a lack of interpretability and physical consistency.

Methodology
This research proposes a knowledge-guided deep learning approach that integrates the physical principles of the Water Cloud Model (WCM) into a Long Short-Term Memory (LSTM) neural network architecture. The methodology aims to leverage the LSTM's capacity to capture spatiotemporal dependencies while enforcing physical constraints.

• Data Inputs: The model utilizes Sentinel-1 SAR data (VH polarization, Ground Range Detected products) and isolates soil backscatter coefficients from total backscatter using WCM principles. It incorporates Landsat-resolution vegetation information (NDVI, albedo) derived via spatiotemporal fusion of MODIS and Landsat data, alongside auxiliary surface characteristics (soil texture, available water capacity, climate grids).
• Model Architecture:
  • Instead of the standard linear regression used in WCM to relate soil backscatter to SM, the study replaces this with an LSTM function ($f_{LSTM}$).
  • The input to the LSTM includes the time series of isolated soil backscatter ($\sigma_{soil}$) and a vector of auxiliary surface characteristics ($v$).
  • The vegetation backscattering factor ($A$) is treated as a learnable parameter within the network, constrained to positive values via an exponential transformation ($A = e^{\log A}$).
• Loss Function: A modified dual-component loss function is employed to ensure physical consistency:
  1. Soil Component ($Loss_{soil}$): Minimizes the mean squared error between the LSTM-predicted SM and in-situ observations.
  2. Boundary Regularization ($Loss_{boundary}$): Enforces physical boundaries (SM must be between 0 and 1) using a penalty term scaled by a regularization factor ($\lambda$).

Key Contributions

• Hybrid Framework: The paper presents a preliminary effort to embed WCM knowledge directly into a deep learning architecture, moving beyond purely empirical or purely data-driven paradigms.
• Isolation of Soil Signal: The approach explicitly utilizes soil backscatter coefficients isolated from total backscatter, combined with high-resolution vegetation data, to drive the LSTM.
• Physical Consistency: By modifying the loss function to include boundary conditions and retaining the semi-physical structure of signal decomposition, the model addresses the "black box" nature of standard deep learning while overcoming the over-simplification of traditional WCM.

Results
The proposed approach was evaluated using a four-fold spatial cross-validation against in-situ SM data from the OzNet Hydrological Monitoring Network in the Yanco agricultural region (semi-arid southeastern Australia).

• Baseline Performance (Standard WCM): The traditional WCM approach yielded correlation coefficients ($R$) between 0.26 and 0.34 and Root Mean Square Errors (RMSE) ranging from 0.08 to 0.10 $m^3/m^3$. The results showed substantial dispersion, with predictions clustering in intermediate SM ranges, indicating the model's inability to capture non-linear relationships under varying surface conditions.
• Proposed Approach Performance: The knowledge-guided LSTM significantly improved performance:
  • RMSE: Reduced to a range of 0.06–0.08 $m^3/m^3$ (a reduction in uncertainty of approximately 0.02 $m^3/m^3$).
  • Correlation: $R$ values improved to between 0.40 and 0.64, with the highest performance observed in the fourth iteration.
  • Stability: The learned vegetation backscattering parameter ($A$) remained stable across iterations (0.0177–0.0201), suggesting consistent characterization of vegetation-backscatter relationships.
  • Limitations: The model exhibited a tendency to slightly underestimate SM at higher values (>0.3 $m^3/m^3$) in some iterations, though the fourth iteration showed improved performance across the full SM range.

Significance and Claims
The authors claim that this research demonstrates the potential of hybrid architectures to address the over-simplification inherent in traditional WCM parameterization, particularly in heterogeneous agricultural landscapes. By integrating physical constraints into the learning process, the approach achieves improved spatial transferability and reduced retrieval uncertainty compared to the standard WCM. The paper positions these findings as preliminary efforts, noting that future work is required to assess the operational capability of this approach across diverse spatial contexts.