[Paper Review] Automatic Estimation of Ice Bottom Surfaces from Radar Imagery

Mingze Xu, David Crandall+2 more|arXiv (Cornell University)|Dec 21, 2017

Cryospheric studies and observations11 references9 citations

TL;DR

This paper proposes a computer vision-based method for automatically reconstructing 3D ice-bottom surfaces from noisy airborne radar imagery using probabilistic graphical models and discrete energy minimization. By combining radar echograms with prior constraints and iterative message passing, the approach achieves significant accuracy improvements over 2D baselines, with 63.9% of pixels within 5 pixels of ground truth in 3D reconstructions of Canadian Arctic ice caps.

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.

Motivation & Objective

• To address the bottleneck of manual, time-intensive labeling of ice-bottom surfaces from radar data in glaciology.
• To enable large-scale, automated 3D reconstruction of subsurface ice structures from airborne radar data.
• To improve accuracy and efficiency over existing 2D layer-finding methods by incorporating 3D spatial constraints and multi-source evidence.
• To develop a scalable, automated pipeline suitable for continental-scale monitoring of polar ice sheets.

Proposed method

• The method formulates 3D ice-bottom surface estimation as a discrete energy minimization problem on a Markov Random Field (MRF).
• It begins with a seed surface generated from layer boundary detection in radar echograms, using a probabilistic template model and smoothness priors.
• Additional evidence sources—such as ice masks, digital elevation models, and optional human feedback—are incorporated via unary and pairwise potentials in the energy function.
• The algorithm uses Tree-Reweighted Message Passing (TRW) for inference, with linear-time generalized distance transforms to accelerate convergence.
• The energy function includes log-Gaussian pairwise potentials to model spatial consistency across adjacent radar slices.
• The final surface is computed by assigning each pixel the label minimizing the sum of messages from neighbors and unary costs.

Experimental results

Research questions

• RQ1 Can a probabilistic graphical model effectively reconstruct 3D ice-bottom surfaces from noisy, multi-beam radar data?
• RQ2 How does incorporating 3D spatial consistency and multi-source evidence improve reconstruction accuracy over 2D methods?
• RQ3 To what extent can automated inference outperform human-labeled ground truth when accounting for labeling noise?
• RQ4 Can message-passing inference on an MRF scale efficiently to continental-scale radar datasets?

Key findings

• The proposed method achieved a 63.9% pixel accuracy within 5 pixels of human-annotated ground truth across 7 sequences of Canadian Arctic ice caps.
• The median mean deviation was 12.2 pixels, and the mean deviation was 11.9 pixels, indicating consistent performance across the dataset.
• The method significantly outperformed three 2D baselines—Viterbi, MCMC, and Dynamic Viterbi—in 3D surface reconstruction.
• When applied to 2D echograms, the method also outperformed the original implementations of Crandall et al. [10] and Lee et al. [11], demonstrating generalization to 2D problems.
• The use of TRW with distance transforms reduced running time to O(lϕρL), enabling efficient inference on large radar sequences.
• The integration of external evidence (e.g., ice masks, human feedback) was shown to be straightforward and effective via the energy function framework.