Eitel et al. (2025) A global analysis of SAR altimetry signals over different landcover types

Identification

• Journal: International Journal of Applied Earth Observation and Geoinformation
• Year: 2025
• Date: 2025-12-03
• Authors: Maximilian Eitel, Michael Schmitt
• DOI: 10.1016/j.jag.2025.105000

Research Groups

• University of the Bundeswehr Munich, Germany

Short Summary

This study analyzes how Sentinel-3 SAR altimetry waveforms respond to different land cover types and what physical characteristics are encoded in the signal. It demonstrates that a feature-enhanced one-dimensional convolutional neural network (1D-CNN) can effectively extract land cover information from these signals, revealing their sensitivity to surface variations despite large footprints.

Objective

• To analyze how Sentinel-3 (S3) synthetic aperture radar (SAR) altimetry waveforms respond to different surface types and what physical characteristics are encoded in the signal.
• To investigate the extent to which altimetric waveforms may provide consistent class-specific information for land cover classification.

Study Configuration

• Spatial Scale: Global analysis covering all continents except Australia, with specific case studies in the Sahara Desert, northeast of the Caspian Sea, and the Amazon basin. The Sentinel-3 SAR altimeter footprint is approximated as an ellipse with a 1.6 km extent in the across-track direction and 300 m along-track.
• Temporal Scale: Sentinel-3 data acquired from July to August 2020. Models were trained for 30 epochs.

Methodology and Data

• Models used:
  • Feature-enhanced one-dimensional Convolutional Neural Network (1D-CNN), specifically an optimized version of the LucasCNN architecture (LucasCNN-FE).
  • Random Forest (baseline).
  • Dummy Classifier (baseline).
• Data sources:
  • Sentinel-3 (S3) SAR altimetry waveforms: Level-2 "enhanced measurement" netCDF data (land product, 128-bin waveform measurements) obtained from the Copernicus Browser.
  • Land cover reference: ESA WorldCover tiles (10 m resolution).
  • Climate classification: Simplified Köppen-Geiger climate classification.
  • Additional features for the 1D-CNN: Standard deviation, Kurtosis, Backscatter coefficient (σ0), Waveform width, Waveform center of gravity, and Simplified Köppen-Geiger-Class.

Main Results

• The feature-enhanced 1D-CNN achieved the highest Macro-F1 score of 0.57 and an overall accuracy of 0.67, outperforming the Random Forest (Macro-F1: 0.46, OA: 0.56) and Dummy Classifier (Macro-F1: 0.17, OA: 0.30).
• An ablation study confirmed the complementary role of both shape-based (standard deviation, kurtosis, width, center of gravity) and contextual (backscatter coefficient, climate class) features, as their removal negatively impacted performance.
• Tree and Water classes were classified with high accuracy (F1-scores above 0.8), indicating distinct radar signatures.
• Significant misclassification occurred between Grass and Bare/Sparse Vegetation (64.6% of Bare/Sparse Vegetation misclassified as Grass), and between Tree, Crop, and Shrub, suggesting similar radar signatures for these classes.
• Classification performance varied across climate zones, with tropical (Macro-F1: 0.523) and polar (Macro-F1: 0.515) regions performing best, while arid and cold regions posed greater challenges.
• Visual comparisons in the Amazon basin demonstrated that the 1D-CNN more consistently detected forest cover and reliably identified river courses and associated water surfaces compared to the Random Forest.

Contributions

• A global characterization of Sentinel-3 waveform behavior over major surface types, highlighting how vegetation, soil, and water conditions are reflected in the signal.
• A benchmark of a feature-enhanced 1D-CNN against traditional baselines (random forest and dummy), including an ablation study clarifying the role of shape-based and contextual features.
• An error analysis with case studies of systematic confusions (e.g., Grass vs. Bare/Sparse vegetation or Crop vs. Tree), which reveals physical similarities in scattering mechanisms between certain classes.
• Demonstration that land information can be retrieved directly from the altimetry waveform itself without requiring additional datasets, providing a foundation for separating coastal water echoes from adjacent land contributions.

Funding

• Universität der Bundeswehr München

Citation

@article{Eitel2025global,
  author = {Eitel, Maximilian and Schmitt, Michael},
  title = {A global analysis of SAR altimetry signals over different landcover types},
  journal = {International Journal of Applied Earth Observation and Geoinformation},
  year = {2025},
  doi = {10.1016/j.jag.2025.105000},
  url = {https://doi.org/10.1016/j.jag.2025.105000}
}