Jiang et al. (2026) Deep-learning full-waveform inversion of snowpack GPR: joint permittivity–resistivity imaging for snow–soil hydrological mapping

Identification

• Journal: Journal of Hydrology
• Year: 2026
• Date: 2026-03-24
• Authors: Yuanjun Jiang, Zohaib Akbar, Ryan Webb, Zhao Binbin, Aftab Anwar, M.M Rehman, M.Z. Mirza
• DOI: 10.1016/j.jhydrol.2026.135374

Research Groups

• Key Laboratory of Mountain Hazards and Engineering Resilience / Institute of Mountain Hazards and Environment, Chinese Academy of Sciences, Chengdu 610213, China
• University of Chinese Academy of Sciences, Beijing 100049, China
• University of Wyoming, Laramie, WY 82071, USA
• State Grid Electric Power Engineering Research Institute Co., Ltd., Beijing 100069, China

Short Summary

This study introduces a hybrid deep learning framework (ViT-BiLSTM) for dual-parameter full-waveform inversion of GPR data, enabling fast and accurate joint imaging of snowpack permittivity and resistivity for hydrological mapping. The framework demonstrates robust performance on synthetic and real-world data, providing spatially coherent snow liquid water content and soil moisture estimates.

Objective

• To introduce a hybrid deep learning framework that combines vision transformers (ViT) with bidirectional long short-term memory (BiLSTM) networks for dual-parameter inversion of permittivity and log-resistivity from GPR data, addressing the limitations of traditional FWI and standard regression-based deep learning in heterogeneous snowpacks.

Study Configuration

• Spatial Scale: 2D synthetic datasets; real-world field GPR data.
• Temporal Scale: Instantaneous imaging of snowpack and underlying soil conditions.

Methodology and Data

• Models used: Hybrid deep learning framework (Vision Transformers (ViT) + Bidirectional Long Short-Term Memory (BiLSTM) networks); gprMax (for synthetic data generation using Finite-Difference Time-Domain (FDTD) simulations).
• Data sources: Synthetic datasets (generated using gprMax for dry, moist, and wet snowpack conditions with underlying soil layers); real-world field GPR data; in situ snowpit observations (for validation).

Main Results

• Robust performance on synthetic 2D datasets:
  • Permittivity: R² = 0.984, SSIM = 0.97, RMSE = 0.066.
  • Log-resistivity: R² = 0.966, SSIM = 0.94, RMSE = 0.086.
• Application to real-world field GPR data produced spatially coherent snow and soil moisture maps.
• Estimated snow liquid water content (LWC) of 2%–4%.
• Estimated soil moisture of 15%–26%.
• Results showed strong agreement with in situ snowpit observations.

Contributions

• Introduction of a novel hybrid deep learning framework (ViT-BiLSTM) for GPR full-waveform inversion, overcoming computational expense and initial condition sensitivity of traditional FWI.
• Enables fast, physically consistent dual-parameter (permittivity and resistivity) imaging, crucial for reliable snowpack interpretation.
• Addresses generalization issues of standard regression-based deep learning in laterally heterogeneous snowpacks by combining global waveform dependencies (ViT) with temporal continuity (BiLSTM).
• Provides a scalable solution for operational snow hydrology, meltwater prediction, and cryospheric monitoring.

Funding

• [Funding information not available in the provided text.]

Citation

@article{Jiang2026Deeplearning,
  author = {Jiang, Yuanjun and Akbar, Zohaib and Webb, Ryan and Binbin, Zhao and Anwar, Aftab and Rehman, M.M and Mirza, M.Z.},
  title = {Deep-learning full-waveform inversion of snowpack GPR: joint permittivity–resistivity imaging for snow–soil hydrological mapping},
  journal = {Journal of Hydrology},
  year = {2026},
  doi = {10.1016/j.jhydrol.2026.135374},
  url = {https://doi.org/10.1016/j.jhydrol.2026.135374}
}