Learning Enhanced Structural Representations with Block-Based Uncertainties for Ocean Floor Mapping

This paper introduces a novel uncertainty-aware framework that combines Vector Quantized Variational Autoencoders (VQ-VAE) with block-based conformal prediction to generate high-resolution, structurally consistent ocean floor maps with spatially adaptive uncertainty estimates, thereby enhancing the reliability of bathymetric data for climate modeling and coastal hazard assessment.

The Gist

Imagine trying to draw a detailed map of the ocean floor, but you only have a blurry, low-resolution satellite photo. You know the general shape of the mountains and valleys underwater, but the fine details—the sharp ridges, the deep trenches, and the hidden canyons—are lost in the fuzziness.

This is the problem scientists face with bathymetry (mapping the ocean floor). Current maps are often too blurry to be useful for predicting tsunamis, storm surges, or climate change impacts. If you smooth out the details too much, you might underestimate how high a tsunami wave will get when it hits the shore.

This paper presents a new "smart map-maker" that does two things:

1. It fills in the missing details to create a sharp, high-definition map.
2. It tells you exactly how much it trusts each part of that map.

Here is how it works, using some everyday analogies:

1. The Problem: The "Blurry Photo" vs. The "Pixelated Guess"

Think of current ocean maps like an old, grainy photo.

• Old Methods (Interpolation): These are like trying to guess what's in the blurry spots by just stretching the pixels. It makes the image smoother, but it turns sharp cliffs into gentle slopes. It's like trying to draw a jagged mountain range with a soft, round marker; you lose the sharp edges.
• Other AI Methods (Deep Learning): These are like artists who are great at making things look pretty and realistic, but they might invent details that aren't there or miss the specific shape of a trench. They also don't usually say, "I'm not sure about this part."

2. The Solution: The "LEGO Master" (VQ-VAE)

The authors used a special type of AI called a VQ-VAE. Imagine you are trying to rebuild a complex LEGO castle from a blurry photo.

• Instead of trying to mold clay (which is smooth and continuous), this AI has a box of specific LEGO bricks (discrete codebook).
• When it sees a part of the ocean floor, it doesn't guess a random shape; it picks the perfect pre-made brick that fits that specific pattern (like a "canyon brick" or a "ridge brick").
• Why this matters: Because it uses these distinct "bricks," it preserves the sharp, jagged edges of the ocean floor. It doesn't smooth them out. It keeps the structure intact, which is crucial for physics simulations (like calculating how a tsunami wave would crash).

3. The Secret Sauce: The "Confidence Checker" (Block-Based Uncertainty)

This is the paper's biggest innovation. Most AI models just give you an answer. This model gives you an answer plus a confidence score.

Imagine you are a weather forecaster.

• Old AI: "It will rain tomorrow." (No idea how sure they are).
• This New AI: "It will rain tomorrow. I am 99% sure about the city center because I have good data there, but I am only 60% sure about the mountains because the data is patchy."

How it works:
The AI divides the ocean map into a grid of blocks (like a checkerboard).

• The "Easy" Blocks: If a block is in an area where we have good sonar data (like a shallow shelf), the AI knows it's accurate. It draws a tight, confident line around its prediction.
• The "Hard" Blocks: If a block is in a deep, complex trench where data is scarce, the AI admits, "This is tricky." It draws a wider, fuzzier line to say, "The real answer could be anywhere within this range."

This is called Block-Based Uncertainty. It adapts to the local complexity. It doesn't treat the whole ocean the same; it knows when to be confident and when to be cautious.

4. Why This Matters for Climate Change

Why do we care about a better map?

• Tsunamis and Storms: To predict how a tsunami wave will travel, you need to know the exact shape of the ocean floor. If you smooth out a trench, the wave might look slower or smaller than it really is. This new method keeps the trenches sharp, leading to better warnings.
• Saving Lives: By knowing exactly where the AI is "guessing" (high uncertainty) and where it is "certain" (low uncertainty), emergency planners can make better decisions. They know which areas need more physical surveys and which predictions they can trust immediately.

The Bottom Line

The authors built a system that acts like a smart, honest cartographer.

• It uses a "LEGO" approach to keep the ocean's sharp features from getting blurry.
• It uses a "grid of confidence" to tell us exactly how reliable its map is in every single spot.

The result? A map that is not only sharper and more accurate than previous methods but also honest about its own limitations, making it a much safer tool for protecting coastal communities from climate disasters.

Technical Summary

1. Problem Statement

Accurate ocean floor mapping (bathymetry) is critical for numerical simulations of environmental phenomena, including tsunami propagation, storm surges, and sea-level rise. However, current global datasets (e.g., GEBCO) suffer from:

• Coarse Resolution: Global sub-kilometer details would take centuries to survey at current rates.
• Data Heterogeneity: Inputs come from diverse sources (multibeam sonar, satellite altimetry) with varying error characteristics and resolution gaps.
• Structural Loss: Conventional interpolation methods (bilinear, bicubic) and standard deep learning super-resolution techniques (SRCNN, GANs) tend to oversmooth critical morphological features like ridges, canyons, and trenches. This can lead to underestimating tsunami wave heights by up to 70%.
• Lack of Uncertainty Quantification: Existing methods fail to provide spatially adaptive confidence estimates, which are essential for risk assessment in climate modeling.

2. Methodology

The authors propose a novel framework combining a Vector Quantized Variational Autoencoder (VQ-VAE) with a Block-Based Uncertainty Awareness (UA) mechanism.

A. Architecture: VQ-VAE with Residual Attention

Instead of using continuous latent representations (which cause oversmoothing), the model employs a discrete latent space:

• Encoder: Maps input bathymetry $x$ to a latent space $z$, which is then quantized ($z_q$) using a learned codebook $C$. This discrete quantization preserves sharp discontinuities and unique morphological features.
• Decoder: Reconstructs high-resolution bathymetry $\hat{x}$ from the quantized latent representation.
• Residual Attention: Integrated to maintain long-range spatial dependencies (e.g., continental shelf margins) and structural consistency.

B. Block-Based Uncertainty Awareness (UA)

To address spatially varying data quality, the authors introduce a modular uncertainty module:

• Spatial Partitioning: The bathymetric map is divided into non-overlapping blocks $B = \{b_1, ..., b_N\}$.
• Error Tracking: During training, the model tracks reconstruction errors for each block using Exponential Moving Averages (EMA).
• Adaptive Weighting: The uncertainty estimate $U_i$ for a block is calculated based on its local reconstruction error relative to historical statistics.
• Loss Function: The training objective incorporates uncertainty-weighted reconstruction loss:
  $$L = \sum_i U_i \cdot |D(z_q)_i - x_i|^2 + \lambda_s(1 - \text{SSIM}) + \lambda_c L_{vq} + \lambda_d L_{div}$$
  • Blocks with higher uncertainty (complex areas or poor data sources) receive higher weights, forcing the model to focus on reducing error in those specific regions.
  • The framework is modular and was also applied to create UA-SRCNN and UA-ESRGAN variants for comparison.

3. Key Contributions

1. VQ-VAE for Bathymetry: Demonstrated that discrete latent representations (via vector quantization) are superior to continuous ones for preserving the sharp structural boundaries of the ocean floor, which is vital for physical modeling.
2. Block-Based Uncertainty Mechanism: Introduced a practical, block-wise uncertainty estimation method that calibrates confidence bounds spatially. It achieved a calibration error of 0.0138, significantly outperforming alternatives (0.0314–0.0374).
3. Comprehensive Evaluation: Provided a rigorous analysis across six major oceanic regions (e.g., Western Pacific, Eastern Atlantic), proving that the method adapts to local bathymetric complexity and data source reliability.

4. Experimental Results

The model was evaluated on GEBCO data (2015 as Low-Res input, 2023 as High-Res target) across 61,520 training samples and 15,380 validation samples.

Performance Metrics (Overall):

Model	SSIM	PSNR (dB)	MSE	UWidth (Uncertainty Width)	CalErr (Calibration Error)
Bilinear (Baseline)	0.7045	15.85	0.0268	-	-
UA-SRCNN	0.8128	18.76	0.0137	0.2966	0.0314
UA-ESRGAN	0.7582	19.20	0.0123	0.2691	0.0374
UA-VQ-VAE (Proposed)	0.9433	26.88	0.0021	0.1046	0.0138

• Reconstruction Quality: The proposed UA-VQ-VAE achieved a 26.88 dB PSNR and 0.9433 SSIM, significantly outperforming the best deep learning alternative (UA-SRCNN) by 8.12 dB and 16.1% in SSIM, respectively.
• Uncertainty Reliability: The model produced much tighter uncertainty widths (0.1046 vs. ~0.29 for others) and lower calibration errors, indicating highly reliable confidence intervals.
• Regional Adaptability:
  • In the Western Pacific (highly complex terrain), the model achieved 0.9385 SSIM and 27.32 dB PSNR with a calibration error of 0.0117.
  • In the Eastern Atlantic, it maintained high fidelity (0.9419 SSIM) with tight confidence bounds.
• Block Size Analysis: Medium-sized blocks (4×4) offered the best balance between local detail preservation and global context, while larger blocks (64×64) excelled in uncertainty calibration.

5. Significance

This work bridges a critical gap in computational geosciences by providing a method that not only enhances the resolution of ocean floor maps but also quantifies the reliability of those enhancements.

• Climate Modeling: By preserving physical structures (ridges, trenches) and providing calibrated uncertainty, the method improves the accuracy of tsunami and storm surge simulations.
• Resource Efficiency: The block-based approach allows the model to adaptively allocate computational capacity to complex areas without oversmoothing simple regions.
• Generalizability: The modular uncertainty framework can be integrated into other deep learning architectures, offering a pathway for more robust Earth Observation (EO) data processing in climate science.

The authors have made pre-trained models and code available via GitHub and Hugging Face to support further research in climate science and marine applications.