RoughNet: Mapping Arctic Sea Ice Roughness Using Diffusion-Based Super-Resolution of Satellite Imagery

Technical Summary: RoughNet – Mapping Arctic Sea Ice Roughness Using Diffusion-Based Super-Resolution

Problem Statement
Accurate estimation of sea ice surface roughness (SIR) is critical for climate modeling (influencing ice-atmosphere interactions and energy balance) and ensuring the safety of Arctic over-ice travel for indigenous communities. However, existing methods for obtaining high-resolution SIR data are limited. Airborne LiDAR surveys provide precise measurements but are costly, logistically challenging, and lack scalability for frequent, large-scale monitoring. Satellite-based proxy methods (e.g., multi-angle optical or SAR backscatter) are constrained by sensor resolution and sampling geometry, often failing to capture fine-scale topographic variability (0.01–5 m). Furthermore, while deep learning has advanced 3D surface inference, there is a gap in applying generative models to reconstruct meter-scale Arctic surface roughness directly from optical satellite imagery, particularly in winter conditions characterized by low contrast and visual homogeneity.

Methodology
The authors propose RoughNet, a conditional diffusion framework designed to map multispectral optical satellite imagery to high-resolution surface elevation residuals.

• Data Sources: The model is trained on paired data from three Arctic regions in Northern Canada (Pond Inlet, Tuktoyaktuk, and Cambridge Bay).
  • Target: Airborne LiDAR data collected during the 2024 IceBird campaigns, processed into 1 m resolution Digital Elevation Models (DEMs). To focus on roughness rather than absolute elevation, these are converted to locally demeaned residual fields by fitting and subtracting a quadratic surface.
  • Conditioning: Sentinel-2 Level-2A multispectral imagery (four 10 m bands: Blue, Green, Red, NIR). The model utilizes stacks of six temporal scenes per region to provide complementary cues regarding illumination and shadows.
• Model Architecture: RoughNet employs a U-Net backbone within a conditional diffusion framework.
  • Input: The network takes a noisy target LiDAR patch (1 m resolution, 256×256 pixels) concatenated with a fused conditioning map derived from six Sentinel-2 patches (resized to 256×256).
  • Fusion Strategy: The six Sentinel-2 views are aggregated into a single attribute-aware map using a shared CNN encoder, a small MLP for metadata (viewing geometry, cloud cover), and a temperature-controlled softmax to weight the most informative views.
  • Training Objective: The model learns to reverse a Gaussian noise corruption process, predicting the clean residual topography directly. The loss function combines per-pixel Mean Squared Error (MSE) with L2 gradient regularization on spatial derivatives to enforce structural fidelity and edge preservation.
• Sampling Strategy: An ablation study compared various noise schedules (Linear vs. Cosine) and samplers (DDPM, DDIM, PLMS). The Cosine PLMS configuration was selected as optimal, balancing low error in roughness amplitude with high statistical fidelity.

Key Contributions

1. First Generative Elevation Synthesis in the Arctic: This work presents the first application of generative diffusion models to synthesize meter-scale sea ice topography directly from optical satellite imagery in a winter Arctic environment.
2. Super-Resolution from 10 m to 1 m: The framework successfully maps 10 m Sentinel-2 inputs to 1 m locally normalized surface residual fields, bypassing the spatial constraints of traditional proxy methods.
3. Statistical Fidelity over Pixel-Wise Perfection: The approach prioritizes the reconstruction of statistical and structural properties (distribution of heights, spatial continuity) over exact pixel-wise correspondence, acknowledging the stochastic nature of surface roughness.

Results
The model was evaluated on validation regions (Pond Inlet, Tuktoyaktuk) and an unseen test region (Cambridge Bay).

• Accuracy: The best-performing model (Cosine PLMS) achieved an out-of-domain Root Mean Squared Error (RMSE) of 0.09 m (9 cm).
• Statistical Properties:
  • Bias: Near zero, indicating no systematic over- or underestimation.
  • Roughness Amplitude: The normalized RMSE ($\sigma$ error) was approximately 17–18%, indicating the reconstruction error is on the order of the intrinsic variability of the roughness field itself.
  • Distributional Fidelity: The Jensen–Shannon Divergence (JSD) between predicted and ground-truth elevation distributions was low (0.07 validation, 0.08 test), confirming the model preserves the correct probability density functions (PDFs), including the log-normal characteristics of sea ice.
  • Spectral Properties: The model captures overall amplitude and orientation but exhibits some over-smoothing of fine-scale textures, leading to a slight misallocation of high-frequency spectral power (log-PSD RMSE ~1.07).
• Generalization: The model generalized well to the unseen Cambridge Bay region, though visual inspection noted that specific elongated ridge structures were not fully captured, likely due to the locally constrained receptive field of the patch-based approach.
• Ablation Findings: The Cosine noise schedule and PLMS sampler outperformed other configurations, with the cosine schedule preserving high-frequency signal information better during early diffusion steps and PLMS providing higher-order trajectory approximation.

Significance and Claims
The paper claims that RoughNet provides a scalable pathway for high-resolution sea ice mapping in data-sparse environments. By demonstrating that generative diffusion models can recover physically meaningful surface structure from optical imagery alone, the work offers a viable alternative to costly airborne surveys.

The authors emphasize that the model's ability to reconstruct centimeter-scale absolute accuracy while preserving realistic roughness statistics represents a substantive advancement in spatial granularity compared to prior work, which typically operates at meter-to-decameter scales. This capability supports two primary applications:

1. Climate Science: Providing realistic roughness parameterizations for ice-atmosphere interaction models.
2. Arctic Safety: Enabling the generation of high-resolution surface maps to assist indigenous communities in planning safer, more efficient travel routes.

The paper modestly acknowledges limitations, including the reliance on fixed-length conditioning (six scenes), the requirement for daylight optical imagery (unlike radar), and the current restriction to immobile landfast ice. It notes that while the model captures overall structure, fine-scale micro-texture reconstruction remains imperfect, and future work is needed to quantify prediction variance for safety-critical navigation.