Jiang et al. (2026) Dual-Branch Machine Learning framework with decoupled architectures for nonlinear interference mitigation in GNSS-R snow depth estimation

Identification

• Journal: International Journal of Applied Earth Observation and Geoinformation
• Year: 2026
• Date: 2026-01-16
• Authors: Zhihao Jiang, Liang Li, Xiuyun Shi, W.C. Ma, He Wang
• DOI: 10.1016/j.jag.2026.105102

Research Groups

Harbin Engineering University, Heilongjiang, China

Short Summary

This study introduces a Dual-Branch Machine Learning framework with decoupled architectures to mitigate nonlinear interference in Global Navigation Satellite System-Reflectometry (GNSS-R) snow depth estimation. The framework significantly improves accuracy, achieving root mean square errors (RMSE) as low as 0.0189 meters and correlation coefficients up to 0.999.

Objective

• To develop a Dual-Branch Machine Learning framework with decoupled architectures to mitigate nonlinear interference and residual errors in GNSS-R snow depth estimation, particularly during snow-free periods, by explicitly separating snow state detection (classification) from snow depth regression.

Study Configuration

• Spatial Scale:
  • P350 station (43.53° N, 114.86° W; 2388.3 meters elevation) and nearby Couch Summit SNOTEL station (43.52° N, 114.80° W; 2068.1 meters elevation), approximately 0.32 kilometers apart.
  • AB33 station (67.25101° N, 150.17254° W; 334.761 meters elevation) in Coldfoot, Alaska, and nearby Coldfoot SNOTEL station (67.25° N, 150.18° W; 326.136 meters elevation), approximately 0.836 kilometers apart.
  • GNSS antenna height: 2 meters.
  • Elevation angle range: P350: 5°–20°; AB33: 5°–35°.
  • Azimuth angle range: 0°–360°.
• Temporal Scale:
  • Data from 2023 (80% for training, 20% for validation) and 2024 (unseen test set).
  • Daily snow depth estimates.
  • GPS L1 band Signal-to-Noise Ratio (SNR) data recorded at 15-second intervals.

Methodology and Data

• Models used:
  • Dual-Branch ML Framework:
    • Snow State Detection (Classification): Multi-Layer Perceptron (MLP), Support Vector Machines (SVM), Random Forests (RF).
    • Snow Depth Estimation (Regression): 1D Convolutional Neural Network (CNN), Support Vector Regression (SVR), Random Forests (RF).
  • Traditional Method: Lomb–Scargle Periodogram (LSP) for frequency extraction, Least Square Fitting (LSF) for SNR trend removal.
  • Comparison Models (Direct Regression): Direct-CNN, Direct-RF, Direct-SVR (without prior classification step).
• Data sources:
  • GNSS-R: GPS L1 band Signal-to-Noise Ratio (SNR) data from P350 and AB33 stations of the Plate Boundary Observation (PBO) network.
  • In-situ: Snow depth data from Couch Summit SNOTEL station (for P350) and Coldfoot SNOTEL station (for AB33).
  • Features extracted from SNR: frequency, amplitude, phase of multipath oscillation, previous day’s snow depth, satellite ID, and azimuth angle.

Main Results

• The proposed Dual-Branch ML framework significantly improved snow depth estimation accuracy:
  • For P350 station, the MLP-CNN method achieved an RMSE of 0.0541 meters and a correlation coefficient of 0.995. This represents a 69.69% reduction in RMSE compared to the traditional method (0.1696 meters RMSE, 0.905 correlation).
  • For AB33 station, the SVM-SVR method achieved an RMSE of 0.0189 meters and a correlation coefficient of 0.999. This represents an 85.96% reduction in RMSE compared to the traditional method (0.1346 meters RMSE, 0.888 correlation).
• The snow state detection module (MLP, SVM, RF) achieved an accuracy higher than 99% at both stations, effectively classifying snow-free and snow-covered periods.
• The decoupled architecture effectively mitigated nonlinear interference, reducing the maximum absolute error from 1.4732 meters (traditional method at P350) to 0.3429 meters (MLP-CNN at P350).
• During snow-free periods, the proposed methods reduced RMSE to nearly zero (e.g., 0.0105 meters for MLP-CNN at P350, 0.0035 meters for SVM-SVR at AB33), while direct regression models (without snow state detection) still showed significant errors (e.g., 0.1312 meters for Direct-CNN at P350, 0.1187 meters for Direct-CNN at AB33).
• MLP-CNN performed best in deep snowpack conditions (P350), while SVM-SVR showed superior accuracy in moderate snow depths (AB33).

Contributions

• Introduction of a novel Dual-Branch Machine Learning framework with decoupled architectures for GNSS-R snow depth estimation, explicitly separating snow state detection from snow depth regression.
• Effective mitigation of nonlinear interference and residual errors, particularly during challenging snow-free periods, providing a robust solution where traditional and direct ML methods often fail.
• Development of an adaptive and automated alternative to manual quality control (QC) parameters, enhancing the reliability and operational applicability of GNSS-R snow depth retrieval.
• Significant enhancement of estimation accuracy, achieving substantially lower RMSE values and higher correlation coefficients compared to existing linear and direct machine learning approaches.
• Establishment of a robust, accurate, and fully automated pipeline for operational snow depth monitoring using GNSS-R.

Funding

• National Key Research and Development Program (No. 2021YFB3901300)
• National Natural Science Foundation of China (No. 62373117, 62403158)

Citation

@article{Jiang2026DualBranch,
  author = {Jiang, Zhihao and Li, Liang and Shi, Xiuyun and Ma, W.C. and Wang, He},
  title = {Dual-Branch Machine Learning framework with decoupled architectures for nonlinear interference mitigation in GNSS-R snow depth estimation},
  journal = {International Journal of Applied Earth Observation and Geoinformation},
  year = {2026},
  doi = {10.1016/j.jag.2026.105102},
  url = {https://doi.org/10.1016/j.jag.2026.105102}
}