Seasonal inundation monitoring and vegetation pattern mapping of the Erguna floodplain by means of a RADARSAT-2 fully polarimetric time series

Lingli Zhao, Jie Yang, Pingxiang Li, Liangpei ZhangRemote 2014 · DOI

TL;DR

This paper leverages RADARSAT-2 fully polarimetric SAR data to map seasonal inundation and vegetation patterns in the Erguna floodplain. By analyzing multi-temporal polarimetric features, the study effectively distinguishes between phenology-driven and flood-driven land cover changes, offering a robust method for monitoring dynamic wetland ecosystems.

What Problem Does This Paper Address?

Floodplains, especially transboundary ones like the Erguna, are critical ecosystems facing increasing pressure from climate change and human activities. Understanding their dynamic hydrological processes and vegetation patterns is crucial for conservation, flood mitigation, and sustainable land management. However, traditional field surveys are impractical for these large, remote, and rapidly changing environments. Optical remote sensing is often hampered by cloud cover, particularly during the crucial rainy season. Synthetic Aperture Radar (SAR) offers a weather-independent solution, and fully polarimetric SAR (PolSAR) provides rich information about scattering mechanisms, making it ideal for characterizing complex wetland environments. This paper addresses the need for a detailed, spatio-temporal understanding of floodplain dynamics, specifically disentangling the influences of vegetation phenology and inundation.

Methods

The study employs a multi-temporal approach using six RADARSAT-2 fully polarimetric SAR images acquired over the Erguna floodplain. Key methodological components include:

  • Polarimetric Feature Extraction: A suite of polarimetric observables were derived from the SAR data, including backscattering coefficients (HH, HV, VV, span), copolarized ratio ($\gamma$), copolarized correlation coefficient ($P_{HHVV}$), copolarized phase difference (CPD), Cloude-Pottier decomposition parameters (Entropy H, Alpha $\alpha$, Anisotropy A), and Yamaguchi decomposition components (odd, double, volume, helix scattering).
  • Absolute Change Estimator (ACE): This technique was used to quantify temporal variations between image pairs, aiding in feature reduction and multi-temporal analysis.
  • Random Forest (RF) Classifier: A robust ensemble learning algorithm was used for both single-temporal and multi-temporal land-cover classification. RF was chosen for its ability to handle high-dimensional data and its tolerance to noise.
  • Trajectory Analysis: To disentangle phenology-driven and flooding-induced changes, land-cover change maps were aggregated into three generic classes: "unchanged," "phenology-induced," and "flooding-induced." Landscape metrics (PLAND, NLSI, IJI) were then calculated on these aggregated maps to analyze spatio-temporal patterns.
  • Hydroperiod Mapping: Open water areas identified in single-temporal classifications were used to generate a hydroperiod map, indicating the probability of inundation.

Datasets

  • RADARSAT-2 Fully Polarimetric SAR Images: Six images were acquired with the same configuration (FQ18, incidence angle 37.4°-38.9°) between September 2012 and August 2013.
  • Ancillary Data:
    • SRTM 90m Digital Elevation Model (DEM) for geometric correction.
  • LiDAR data (0.5m resolution) for DEM, DSM, and CHM generation, used for ROI selection.
  • Aerial CCD images for ROI selection.
  • Ground truth data (field campaigns, photos, vegetation height/weight measurements) collected between September 2012 and August 2013 for training and validation.
  • Meteorological data (precipitation and temperature) from the China Meteorological Data Sharing Service System.

Key Insights and Ideas

  • Disentangling Phenology and Flooding: The study successfully demonstrates a method to differentiate between land-cover changes caused by vegetation phenology (e.g., growth, senescence) and those caused by seasonal inundation. This is crucial for accurate ecological monitoring.
  • Superiority of Multi-Temporal PolSAR: Multi-temporal analysis significantly improves land-cover classification accuracy and provides a more stable representation of vegetation patterns compared to single-temporal classifications, which are highly variable due to dynamic environmental conditions.
  • HV Polarization Dominance: The HV polarization intensity was identified as the most important feature for land-cover classification in this floodplain environment, highlighting its sensitivity to vegetation structure and inundation.
  • Value of Quad-Polarimetry: While intensity features are paramount, composite polarimetric features, particularly volume scattering, offer valuable complementary information for understanding complex scattering mechanisms in wetlands.
  • Flooding Dynamics: The hydroperiod map effectively visualizes inundation probabilities, revealing that meander scars and low-lying areas are most susceptible to flooding, while forests along riverbanks may act as buffers.
  • Trajectory Analysis for Change Drivers: Aggregating classification results and analyzing landscape metrics (PLAND, NLSI, IJI) provides a powerful way to track and interpret the spatio-temporal impact of phenology and flooding.

Key Results

  • Classification Accuracy: Multi-temporal RF classification achieved high accuracies, with an overall accuracy (OA) of 99.31% and a kappa index of 0.991. Single-temporal classifications showed more variability, with the highest OA of 98.10% for July 10, 2013, and the lowest of 89.11% for August 27, 2013.
  • Feature Importance: HV intensity was consistently the most important feature. Among composite features, volume scattering showed high importance, followed by Entropy (H) and the copolarized ratio ($\gamma$).
  • Phenology vs. Flooding: Phenology-induced changes were more prominent in the early part of the study period, while flooding became the dominant driver of change during the rainy season, significantly impacting meadow and bog areas.
  • Hydroperiod Map: The map clearly delineated areas with high inundation probability, particularly along river channels and in low-lying meadows.

Limitations and Future Work

  • Inundation Time Uncertainty: The hydroperiod map indicates the probability of flooding but cannot ascertain the precise inundation time for each area from the six images.
  • Forest Inundation Detection: Inundation under dense forest canopy remains challenging to detect with the current classification technique.
  • Human-Induced Changes: The study acknowledges that human-induced changes (land reclamation, tourism) are impacting the floodplain but focuses primarily on phenology and flooding. Future work is suggested to incorporate these anthropogenic factors into long-term series analysis.
  • Future Missions: The authors highlight the potential of upcoming missions like Sentinel-1 and Radarsat Constellation Mission for broader and more frequent floodplain monitoring.

Summary updated

BibTeX
@inproceedings{zhao2014seasonal,
  author = {Lingli Zhao and Jie Yang and Pingxiang Li and Liangpei Zhang},
  title = {{Seasonal inundation monitoring and vegetation pattern mapping of the Erguna floodplain by means of a RADARSAT-2 fully polarimetric time series}},
  booktitle = {Remote},
  year = {2014},
  doi = {10.1016/j.rse.2014.06.026},
}
Figure