Automatic estimation of ice bottom surfaces from radar imagery

Mingze Xu, David J. Crandall, Geoffrey C. Fox, John D. Paden

TL;DR

This paper introduces a novel computer vision approach to automatically reconstruct the 3D subsurface structure of polar ice sheets from noisy radar data. By framing the problem as a discrete energy minimization task on a graphical model, the authors achieve significantly more accurate and efficient ice-bottom surface estimation compared to existing methods, paving the way for large-scale glaciological analysis.

What Problem Does This Paper Address?

Understanding and predicting the impact of global climate change hinges on accurate data about polar ice sheets. Ground-penetrating radar systems can collect vast amounts of subsurface data, but manually interpreting this noisy, complex information at a continental scale is impractical. Existing automated methods primarily focus on 2D echograms (radar cross-sections) along the radar's flight path, which are often less accurate than human interpretations and don't capture the broader 3D structure. This paper addresses the critical need for an automated, 3D reconstruction technique that can handle noisy radar data and provide a more comprehensive view of the ice sheet's subsurface.

Methods

The core of the proposed method is a computer vision-based technique that formulates the 3D ice-bottom surface estimation as a discrete energy minimization problem on a Markov Random Field (MRF).

  • Graphical Model: The ice-bottom surface is parameterized as a sequence of slices, and the goal is to infer the optimal surface configuration that minimizes a defined energy function.
  • Energy Function: This function comprises two main components:
  • Unary Term ($\psi_1$): This measures how well a potential surface label agrees with the observed radar data. It incorporates:
  • A template model learned from labeled data to capture the expected vertical profile of the ice-bottom surface.
  • A "ice-air boundary" constraint to ensure the ice-bottom surface is below the ice-air surface.
  • A "bottom bin" constraint using weak evidence from the radar's bottom bin to further refine the surface.
  • Pairwise Term ($\psi_2$): This encourages smoothness and continuity of the estimated surface by penalizing differences between adjacent surface labels, both within the same slice (columns) and across adjacent slices (rows).
  • Inference Algorithm: Due to the large size of the MRF, the authors employ Sequential Tree-Reweighted Message Passing (TRW). This algorithm breaks the MRF into monotonic chains and performs belief propagation, offering faster convergence and reduced memory usage compared to traditional message passing methods. They also utilize linear-time generalized distance transforms to optimize the TRW process.

Datasets

The authors evaluated their algorithm on the basal topography of the Canadian Arctic Archipelago (CAA) ice caps.

  • Data Source: Data was collected by the Multichannel Coherent Radar Depth Sounder (MCORDS) instrument.
  • Dataset Size: The evaluation used 7 topographic sequences, each containing over 3000 radar images, covering approximately 50km of flight data per sequence.
  • Ground Truth: Human-annotated ice-air surface data was used for training the template model and binary costs, and for evaluating the performance of the algorithm. The authors acknowledge that human labels can have inaccuracies due to noise and inherent difficulty in precise boundary tracking.

Key Insights and Ideas

  • 3D Reconstruction from 2D Data: The paper demonstrates that by cleverly combining evidence from multiple radar beams (nadir, left, right) and leveraging graphical models, it's possible to move beyond 2D cross-section analysis to achieve a true 3D reconstruction of the ice sheet's subsurface.
  • Probabilistic Graphical Models for Noisy Data: The MRF framework is well-suited for integrating diverse, noisy sources of evidence (radar signals, prior knowledge, constraints) to infer a globally consistent and smooth surface.
  • TRW for Efficient Inference: The choice of TRW as the inference algorithm is crucial for handling the large-scale MRF, enabling practical application of the method to extensive radar datasets.
  • Weak Constraints Matter: Even seemingly weak constraints, like the "bottom bin" evidence, can significantly improve the accuracy of the surface estimation by helping to disambiguate boundaries in noisy data.

Key Results

  • Improved Accuracy: The proposed method significantly outperforms existing baselines, including adaptations of 2D layer-finding techniques and a Markov Chain Monte Carlo (MCMC) approach.
  • For ice-bottom surfaces, the mean error was approximately 11.9 pixels, with a median-of-means error of 12.2 pixels.
  • The percentage of correctly labeled pixels was 35.9%, rising to 63.9% within a 5-pixel tolerance, which is considered a meaningful statistic given the data noise.
  • Bedrock Layering: The technique also showed strong performance in estimating bedrock layers, with a mean error of 4.1 pixels and 81.4% correct labels within 5 pixels.
  • Efficiency: The use of TRW and generalized distance transforms leads to an efficient inference process.

Limitations and Future Work

  • Human Label Inaccuracies: The authors acknowledge that the ground truth labels themselves can be imperfect due to noise and the inherent difficulty of precise boundary identification, which can affect evaluation metrics.
  • Future Work: The paper suggests that additional evidence, such as ice masks or human feedback on specific pixels, can be easily incorporated into the energy function to further refine the estimates. This opens avenues for interactive or semi-automated refinement of the 3D reconstructions.

Summary updated

Abstract

Ground-penetrating radar on planes and satellites now makes it practical to collect 3D observations of the subsurface structure of the polar ice sheets, providing crucial data for understanding and tracking global climate change. But converting these noisy readings into useful observations is generally done by hand, which is impractical at a continental scale. In this paper, we propose a computer vision-based technique for extracting 3D ice-bottom surfaces by viewing the task as an inference problem on a probabilistic graphical model. We first generate a seed surface subject to a set of constraints, and then incorporate additional sources of evidence to refine it via discrete energy minimization. We evaluate the performance of the tracking algorithm on 7 topographic sequences (each with over 3000 radar images) collected from the Canadian Arctic Archipelago with respect to human-labeled ground truth.
BibTeX
@inproceedings{xu2017automatic,
  author = {Mingze Xu and David J. Crandall and Geoffrey C. Fox and John D. Paden},
  title = {{Automatic estimation of ice bottom surfaces from radar imagery}},
  booktitle = {ICIP},
  year = {2017},
  doi = {10.1109/icip.2017.8296299},
}
Figure