Lee et al. (2026) ESCAPE: An ensemble-based self-calibrated autoencoder with physics-informed estimation of high-resolution soil moisture and surface roughness from ALOS-2/PALSAR-2 polarimetric observations

Identification

• Journal: International Journal of Applied Earth Observation and Geoinformation
• Year: 2026
• Date: 2026-02-26
• Authors: Jaehoon Lee, Haemi Park, Jungho Im, Christian N. Koyama, Rogelio Ruzcko Tobias, Takeo Tadono
• DOI: 10.1016/j.jag.2026.105206

Research Groups

• Department Environment and Energy Engineering, Gwangju Institute of Science and Technology (GIST), Gwangju, The Republic of Korea
• Global Environmental Studies, Sophia University, Tokyo, Japan
• Department of Civil, Urban, Earth, and Environmental Engineering, Ulsan National Institute of Science and Technology (UNIST), Ulsan, The Republic of Korea
• Graduate School of Carbon Neutrality, UNIST, Ulsan, The Republic of Korea
• Graduate School of Artificial Intelligence, UNIST, Ulsan, The Republic of Korea
• Earth Observation Research Center, Japan Aerospace Exploration Agency (JAXA), Ibaraki, Japan

Short Summary

This study introduces ESCAPE, an ensemble-based self-calibrated autoencoder with physics-informed estimation, to retrieve high-resolution soil moisture (SM) and surface roughness (Hrms) from ALOS-2/PALSAR-2 polarimetric observations without requiring in-situ SM for direct training, demonstrating robust performance and improved generalization across diverse environments.

Objective

• Can physics-informed neural networks achieve generalizable soil moisture (SM) estimation with a limited number of training samples?
• How does ill-posedness in physical scattering equations influence uncertainty in Physics-Informed Neural Network (PINN)-based SM estimation?
• Can ensembling PINNs with different weight initializations reduce predictive uncertainty and improve SM estimation accuracy?

Study Configuration

• Spatial Scale: Global (ALOS-2/PALSAR-2 observations), with evaluation over three distinct regions: Hydrological Open Air Laboratory (HOAL) in Austria, REMEDHUS network in Spain, and Ibaraki site in Japan. ALOS-2/PALSAR-2 data at approximately 6 m spatial resolution.
• Temporal Scale: ALOS-2/PALSAR-2 observations from 2014 to present (14-day revisit period). In-situ SM data from ISMN (since 2011, various temporal resolutions) and JAXA (daily, 2021–2023). Specific SAR acquisitions for evaluation in May and June 2017.

Methodology and Data

• Models used:
  • ESCAPE (Ensemble-based Self-Calibrated Autoencoder with Physics-informed Estimation)
  • Physics-informed Neural Network (PINN) ensemble (10 members)
  • Coupled Water Cloud Model (WCM) for vegetation scattering
  • Oh2004 surface scattering model for bare soil
• Data sources:
  • Satellite:
    • ALOS-2/PALSAR-2 (L-band, full polarimetric, Level-1.1 Single Look Complex (SLC) data, ~6 m spatial resolution) for backscatter coefficients (σVV, σHH, σVH), Radar Vegetation Index based on Freeman-Durden decomposition (RVIFD), and local incidence angle (θ).
    • Sentinel-2 (for S2GLC land cover product).
  • Observation (In-situ):
    • International Soil Moisture Network (ISMN) for soil moisture (0–5 cm depth) from HOAL (Austria) and REMEDHUS (Spain) networks.
    • JAXA in-situ data from Ibaraki (Japan) for soil moisture (5 cm depth) and surface root mean square height (Hrms).
  • Reanalysis/Auxiliary:
    • Sentinel-2 Global Land Cover (S2GLC) product (10 m spatial resolution, 2017) as auxiliary land cover input.
    • SRTM 1 sec HGT data for terrain correction.

Main Results

• ESCAPE achieved robust soil moisture (SM) and surface roughness (Hrms) estimation without direct in-situ SM measurements for training.
• Spatial Evaluation (HOAL + REMEDHUS combined):
  • SM: Correlation coefficient (R) = 0.701, unbiased root mean squared difference (ubRMSD) = 0.089 m^3 m^-3, bias = 0.006 m^3 m^-3.
• Temporal Evaluation (Ibaraki, Japan):
  • SM: R = 0.568, ubRMSD = 0.06 m^3 m^-3, bias = -0.13 m^3 m^-3.
  • Hrms: R = 0.793, ubRMSD = 0.202 cm, bias = 0.444 cm.
• The ensemble strategy significantly improved robustness by mitigating uncertainty from ill-posed physical constraints and gradient-based optimization, reducing the influence of poorly converged individual members.
• Reconstructed backscatter (σrec) showed very strong agreement with measured backscatter (R > 0.97 across all polarizations).
• Physically simulated backscatter (σsim) showed moderate agreement with measured backscatter (R ranging from 0.716 to 0.808 across polarizations).
• Sensitivity analysis revealed that simulated backscatter is substantially more sensitive to the WCM parameter B than to A, explaining A's near-constant behavior during optimization.

Contributions

• Proposes ESCAPE, a novel self-calibrated, satellite-only framework for high-resolution estimation of soil moisture and surface roughness from polarimetric SAR observations.
• Introduces an ensemble-based physics-informed autoencoder that effectively addresses the challenges of ill-posed inversion and limited training data by integrating physical scattering models without direct in-situ SM training targets.
• Demonstrates improved generalization capabilities across diverse and heterogeneous environments compared to conventional approaches that rely on extensive in-situ calibration.
• Provides a robust and scalable framework adaptable to current and future polarimetric SAR missions (e.g., Sentinel-1, ALOS-4/PALSAR-3, NISAR, ROSE-L) for geophysical parameter retrieval in data-sparse regions.

Funding

• R&D Program for Forest Science Technology (RS-2025-02213492) provided by Korea Forest Service (Korea Forestry Promotion Institute)
• JAXA Earth Observation Research Announcement (EO-RA3, PI No. ER3A2N204)

Citation

@article{Lee2026ESCAPE,
  author = {Lee, Jaehoon and Park, Haemi and Im, Jungho and Koyama, Christian N. and Tobias, Rogelio Ruzcko and Tadono, Takeo},
  title = {ESCAPE: An ensemble-based self-calibrated autoencoder with physics-informed estimation of high-resolution soil moisture and surface roughness from ALOS-2/PALSAR-2 polarimetric observations},
  journal = {International Journal of Applied Earth Observation and Geoinformation},
  year = {2026},
  doi = {10.1016/j.jag.2026.105206},
  url = {https://doi.org/10.1016/j.jag.2026.105206}
}