From Local Training to Large-Scale Mapping: A Comparative Assessment of Machine Learning and Deep Learning for Transferable Satellite-Derived Bathymetry

The Gist

Imagine the ocean floor as a giant, hidden puzzle. For ships to navigate safely and for scientists to study coral reefs, they need to know exactly how deep the water is. Traditionally, mapping this "underwater landscape" requires expensive boats with sonar or planes with lasers, which is slow and can only cover small areas.

This paper explores a cheaper, faster way: using satellite photos (specifically from the Sentinel-2 satellite) to "see" the depth of the water. It's like trying to guess how deep a swimming pool is just by looking at the color of the water from above. The deeper the water, the darker and bluer it looks, but it's a tricky relationship that changes depending on the sand, the coral, and how sunny it is.

The researchers asked a big question: Can we teach a computer to look at a satellite photo of one reef, learn the rules, and then accurately guess the depth of a completely different reef thousands of miles away?

Here is how they solved it, explained simply:

1. The "Old Way" vs. The "New Way"

The team compared two types of computer learners:

• The "Pixel Counter" (Random Forest): This is like a student who memorizes that "light blue means 2 meters deep" and "dark blue means 10 meters deep" based on specific examples. It works great if you show it the same pool again, but if you take it to a different pool with different sand or lighting, it gets confused.
• The "Pattern Detective" (Deep Learning): These are advanced AI models (like ResNet and ConvNeXt) that don't just look at single pixels. They look at the whole picture, understanding how the water color changes as it slopes down a reef. They are like a student who understands the physics of light and water, not just the colors.

The Result: The "Pattern Detectives" (Deep Learning) were much better at guessing the depth of the new reefs than the "Pixel Counter." While the Pixel Counter failed when moved to a new location, the Deep Learning models kept their cool, though they still made some mistakes.

2. The Secret Ingredient: Don't Chop the Puzzle

One of the most surprising findings was about how they fed the data to the computer.

• The Bad Way (Random Slices): Imagine taking a photo of a coral reef, cutting it into tiny, random squares, and shuffling them. You lose the context. The computer sees a piece of a reef slope but doesn't know it's connected to a lagoon.
• The Good Way (Continuous Blocks): Instead, the researchers kept the reef pieces connected, like keeping a jigsaw puzzle together. They fed the computer large, continuous chunks of the reef.

The Analogy: It's the difference between learning a language by memorizing random words versus reading whole sentences. By keeping the reef "whole," the AI learned the shape of the underwater world, not just the colors. This made the AI much more accurate and better at traveling to new locations.

3. The "Shallow Water" Focus

The researchers realized that for ships, the most dangerous part is the very shallow water (where you might hit a reef). Standard math treats a 1-meter error in deep water the same as a 1-meter error in shallow water. But a 1-meter error in 2 meters of water is a disaster; in 20 meters, it's fine.

They invented a special "Smooth Weight Function" (a fancy way of saying a scoring system). Think of it like a teacher grading a test who gives extra credit for getting the shallow water answers right. This forced the AI to pay extra attention to the dangerous, shallow zones, making those predictions much more precise.

4. The "Time-Lapse" Trick

Satellites pass over the same spot many times. The water might look different on different days because of the sun's angle, clouds, or tides.

• The Strategy: Instead of picking just one photo, the team used 10 different photos of the same reef taken on different days.
• The Result: They took the "middle" (median) of all these guesses. If one photo was cloudy or had a weird reflection, the other photos canceled it out. This made the final map much smoother and more reliable, like taking a long-exposure photo to remove noise.

The Bottom Line

The study found that while we can't yet map the entire ocean floor with perfect, survey-grade accuracy using just satellites, we are getting much closer.

• Deep Learning models are the winners, especially when they are trained on connected chunks of reefs rather than random bits.
• By focusing on shallow water and using multiple days of photos, they achieved a level of accuracy that is "good enough" for many applications, even when moving from one part of the world to another.
• However, moving from one reef to a totally different one still causes some errors (the "transfer gap"). The AI is good, but it's not perfect yet because every ocean has unique secrets (different sand, different water clarity) that are hard to learn without seeing them first.

In short: Don't chop the puzzle, focus on the shallow parts, and look at the picture many times over different days. That's the recipe for the best satellite ocean maps we have today.

Technical Summary

Technical Summary: From Local Training to Large-Scale Mapping

Problem Statement
Satellite-derived bathymetry (SDB) offers a cost-effective alternative to traditional multibeam echosounders and airborne LiDAR for mapping shallow waters. However, existing methods face significant challenges in scalability and cross-regional generalizability, particularly in optically complex coastal environments. Classical approaches, including linear regression and log band-ratio methods, are often site-dependent and struggle to retrieve depths beyond approximately 10 meters. While machine learning (ML) methods like Random Forest (RF) have improved accuracy, they remain sensitive to local environmental variability (e.g., water turbidity, seabed composition). Deep learning (DL) presents a potential solution for learning complex spectral-depth relationships, yet systematic investigations into the transferability of DL models across geographically distinct regions are limited. Furthermore, many prior studies rely on within-region random data splits, which may overestimate generalization capabilities by allowing models to learn site-specific spectral signatures rather than transferable physical relationships.

Methodology
This study systematically evaluates and compares classical ML and modern deep learning architectures for transferable SDB using multispectral Sentinel-2 imagery over a 0–20 m depth range.

• Study Areas and Data: The framework utilizes training data from Pratas Island (South China Sea) and selected reef systems in the Great Barrier Reef (GBR). Ground truth is derived from airborne LiDAR (Pratas) and a 30-m DEM (GBR). To rigorously assess cross-regional generalization, the models are tested on spatially independent hold-out sites: Ashmore Reef and Cartier Reef in northwestern Australia.
• Models: A Random Forest model serves as a baseline. Four deep learning architectures are evaluated within a unified DeepLabV3+ framework using ImageNet-pretrained encoders: ResNet-50, ResNet-101, EfficientNet-B4, and ConvNeXt-Large.
• Training Configurations:
  • Input: Sentinel-2 Level-2A surface reflectance. A logarithmic transformation is applied to DL inputs to align with classical SDB physics, while RF uses raw reflectance.
  • Spatial Continuity: Two data-splitting strategies are compared: random patch splitting (Strategy 1) versus spatially continuous splitting (Strategy 2), where patches are extracted as contiguous reef blocks with 50% overlap to preserve geomorphic structure.
  • Loss Functions: The study introduces a Smooth Weight Function (SWF)-weighted RMSE loss. This function applies a depth-dependent weight ($w_i$) to emphasize shallow-water pixels (critical for navigation) while retaining contributions from intermediate depths, contrasting with standard RMSE and Relative Percentage Error (RPE).
  • Temporal Aggregation: Multi-temporal repeat imagery is used for both training and inference. During inference, predictions from multiple acquisitions are aggregated using the median to reduce noise from sun-glint, atmospheric conditions, and tides.
• Benchmarking: The proposed architectures are also benchmarked against task-specific networks (U-Net, Swin-BathyUNet) using the public MagicBathyNet dataset (aerial RGB imagery) to assess parameter efficiency.

Key Contributions

1. Rigorous Transferability Assessment: Unlike many prior studies using within-region random splits, this work employs spatially independent hold-out regions (Ashmore and Cartier) to provide a more honest evaluation of cross-regional generalization.
2. Spatial Continuity Strategy: The study demonstrates that preserving spatial continuity during training (using contiguous reef blocks) is the most impactful design choice, significantly outperforming random patch splitting in both intra-regional accuracy and cross-regional transferability.
3. Depth-Aware Loss Function: The introduction of the SWF-weighted RMSE loss allows for flexible control over prediction accuracy across the depth range, specifically enhancing performance in shallow waters where optical retrieval is most challenging.
4. Architecture Comparison: The paper provides a comprehensive comparison of general-purpose convolutional backbones, showing that models with fewer parameters can match or exceed the performance of larger, task-specific transformer architectures when trained on unified datasets.

Results

• Intra-Regional Performance: Deep learning models, particularly ResNet-50 and ResNet-101, achieved intra-regional RMSE values between 1.15 m and 1.26 m over the 0–20 m range. In very shallow waters (≤3 m), RMSE dropped to as low as 0.26 m (ResNet-101). While Random Forest performed competitively in intra-regional tests (RMSE 1.53 m), it lacked the structural detail of DL models.
• Cross-Regional Generalization: Deep learning models exhibited greater robustness to geographic shifts than Random Forest.
  • Random Forest accuracy degraded significantly, with RMSE increasing from 1.53 m (intra-regional) to 2.99–3.78 m (cross-regional).
  • Deep learning models showed moderate degradation. The best-performing models (ResNet-50, ResNet-101, ConvNeXt-Large) achieved cross-regional RMSE values of approximately 2.46–2.98 m.
  • ConvNeXt-Large demonstrated the strongest relative robustness, with cross-regional RMSE increasing by only ~42% compared to its intra-regional performance. At intermediate depths (5–11 m), ConvNeXt-Large achieved RMSE values approaching ~1 m at Cartier Reef.
• Loss Function and Data Strategy:
  • The SWF loss consistently improved shallow-water accuracy (≤3 m) compared to standard RMSE and RPE, particularly for ConvNeXt-Large.
  • Spatially continuous training (Strategy 2) reduced intra-regional RMSE by up to 63% compared to random splitting (Strategy 1) for ResNet-50. It also yielded the lowest cross-regional errors across all architectures.
• Benchmarking: On the MagicBathyNet benchmark (0–16 m), the proposed models achieved RMSE values between 0.19 m and 0.22 m, outperforming the Swin-BathyUNet (0.24–0.53 m) and baseline U-Net while using substantially fewer parameters (e.g., 58.7 M vs. 395 M for Swin-BathyUNet).

Significance and Claims
The paper claims to demonstrate a practical pathway toward transferable, large-area SDB using general-purpose deep learning models. The authors assert that successful cross-regional performance relies not merely on architectural complexity but on the alignment of three components: adherence to optical constraints, preservation of geomorphic continuity during training, and depth-aware optimization.

The study acknowledges that while deep learning models have made meaningful progress, they do not yet achieve operational survey-grade accuracy (IHO Order 2) across the full 0–20 m range under cross-regional conditions. The residual error is attributed to environmental variability (substrate composition, water-column optics) and the limitations of purely data-driven regression in the absence of explicit physical priors. The authors position this work as a rigorous baseline using off-the-shelf architectures, suggesting that the next step for achieving truly scalable SDB lies in developing dedicated physics-informed neural networks that explicitly model substrate albedo and water-column attenuation. The release of optimized network architectures and pretrained weights is intended to facilitate further research and application in shallow coastal and reef environments.