[Paper Review] Skip-WaveNet: A Wavelet based Multi-scale Architecture to Trace Snow Layers in Radar Echograms

Debvrat Varshney, Masoud Yari+5 more|arXiv (Cornell University)|Oct 30, 2023

Cryospheric studies and observations4 citations

TL;DR

Skip-WaveNet proposes a wavelet-based multi-scale deep learning architecture to automatically trace snow layers in airborne radar echograms, leveraging discrete wavelet transforms for multi-resolution feature extraction and skip connections for improved gradient flow. It achieves a mean absolute error of 3.31 pixels and 94.3% average precision, outperforming state-of-the-art methods in snow layer detection accuracy and generalizability, with strong agreement to in-situ drilling measurements.

ABSTRACT

Airborne radar sensors capture the profile of snow layers present on top of an ice sheet. Accurate tracking of these layers is essential to calculate their thicknesses, which are required to investigate the contribution of polar ice cap melt to sea-level rise. However, automatically processing the radar echograms to detect the underlying snow layers is a challenging problem. In our work, we develop wavelet-based multi-scale deep learning architectures for these radar echograms to improve snow layer detection. These architectures estimate the layer depths with a mean absolute error of 3.31 pixels and 94.3% average precision, achieving higher generalizability as compared to state-of-the-art snow layer detection networks. These depth estimates also agree well with physically drilled stake measurements. Such robust architectures can be used on echograms from future missions to efficiently trace snow layers, estimate their individual thicknesses and thus support sea-level rise projection models.

Motivation & Objective

• To address the challenge of accurate, automated snow layer detection in radar echograms from airborne ice-penetrating radar.
• To improve generalizability and robustness of deep learning models in detecting snow layers across diverse radar echogram conditions.
• To develop a multi-scale architecture that captures hierarchical features across varying spatial scales in radar data.
• To validate model predictions against physically drilled stake measurements for real-world accuracy.
• To support future climate modeling by enabling efficient, precise estimation of snow layer thicknesses from radar data.

Proposed method

• The architecture employs a wavelet-based encoder-decoder framework using discrete wavelet transforms (DWT) to decompose radar echograms into multi-scale subbands.
• Multi-scale features are extracted from the approximation and detail coefficients at multiple decomposition levels to capture both coarse and fine-grained layer structures.
• Skip connections are integrated between encoder and decoder blocks to preserve spatial resolution and enhance gradient flow during training.
• A U-Net-like skip-connection design enables end-to-end learning of depth estimates with improved feature reconstruction.
• The model is trained using a pixel-wise regression loss to predict layer depths, with optimization via Adam and data augmentation for robustness.
• Wavelet decomposition is applied prior to network input, enabling the model to process multi-scale representations without additional convolutional layers.

Experimental results

Research questions

• RQ1 Can a wavelet-based multi-scale architecture improve the accuracy and generalizability of snow layer detection in radar echograms compared to existing deep learning methods?
• RQ2 How does the integration of skip connections in a wavelet-transformed feature space affect depth estimation performance?
• RQ3 To what extent do the model's predicted layer depths correlate with physically measured snow layer depths from drilling stakes?
• RQ4 Can the proposed architecture effectively handle variations in signal-to-noise ratio and layer continuity across diverse radar echograms?
• RQ5 How does the use of discrete wavelet transforms enhance feature representation for snow layer segmentation in radar data?

Key findings

• The model achieves a mean absolute error (MAE) of 3.31 pixels in estimating snow layer depths, demonstrating high precision in depth regression.
• It attains 94.3% average precision in detecting snow layers, indicating strong performance in localization and segmentation.
• The predictions show strong agreement with in-situ measurements from physically drilled stakes, validating real-world reliability.
• The architecture generalizes better than state-of-the-art networks, particularly in challenging echogram conditions with low signal-to-noise ratios.
• The use of wavelet transforms enables effective multi-scale feature learning without increasing model complexity or parameter count.
• The skip-connection mechanism significantly improves training stability and feature recovery, especially in deep network stacks.