Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

From Bands to Depth: Understanding Bathymetry Decisions on Sentinel-2

Satyaki Roy Chowdhury, Aswathnarayan Radhakrishnan, Hsiao Jou Hsu, Hari Subramoni, Joachim Moortgat

TL;DR

This study tackles Sentinel‑2 derived bathymetry by introducing Swin‑BathyUNet, a transformer‑based U‑Net augmented with decoder‑conditioned cross‑attention to improve depth inference and boundary fidelity in shallow waters. It systematically analyzes spectral contributions via Leave‑One‑Band‑Out (LOBO), explains predictions with A‑CAM‑R and a retention test to verify explanations reflect evidence the model relies on, and demonstrates that cross‑region generalization degrades with depth, guiding practical transfer strategies. Key findings include that Green is the most informative band for depth, that preserving wide receptive fields and robust glint filtering enhances robustness, and that cross‑region transfer benefits from depth‑aware calibration and modest fine‑tuning at light target sites. The work provides concrete, physics‑aligned guidance for data curation and model design to enable reliable SDB deployment across regions, along with multi‑modal data fusion insights and interpretable diagnostics for operational use.

Abstract

Deploying Sentinel-2 satellite derived bathymetry (SDB) robustly across sites remains challenging. We analyze a Swin-Transformer based U-Net model (Swin-BathyUNet) to understand how it infers depth and when its predictions are trustworthy. A leave-one-band out study ranks spectral importance to the different bands consistent with shallow water optics. We adapt ablation-based CAM to regression (A-CAM-R) and validate the reliability via a performance retention test: keeping only the top-p% salient pixels while neutralizing the rest causes large, monotonic RMSE increase, indicating explanations localize on evidence the model relies on. Attention ablations show decoder conditioned cross attention on skips is an effective upgrade, improving robustness to glint/foam. Cross-region inference (train on one site, test on another) reveals depth-dependent degradation: MAE rises nearly linearly with depth, and bimodal depth distributions exacerbate mid/deep errors. Practical guidance follows: maintain wide receptive fields, preserve radiometric fidelity in green/blue channels, pre-filter bright high variance near shore, and pair light target site fine tuning with depth aware calibration to transfer across regions.

From Bands to Depth: Understanding Bathymetry Decisions on Sentinel-2

TL;DR

This study tackles Sentinel‑2 derived bathymetry by introducing Swin‑BathyUNet, a transformer‑based U‑Net augmented with decoder‑conditioned cross‑attention to improve depth inference and boundary fidelity in shallow waters. It systematically analyzes spectral contributions via Leave‑One‑Band‑Out (LOBO), explains predictions with A‑CAM‑R and a retention test to verify explanations reflect evidence the model relies on, and demonstrates that cross‑region generalization degrades with depth, guiding practical transfer strategies. Key findings include that Green is the most informative band for depth, that preserving wide receptive fields and robust glint filtering enhances robustness, and that cross‑region transfer benefits from depth‑aware calibration and modest fine‑tuning at light target sites. The work provides concrete, physics‑aligned guidance for data curation and model design to enable reliable SDB deployment across regions, along with multi‑modal data fusion insights and interpretable diagnostics for operational use.

Abstract

Deploying Sentinel-2 satellite derived bathymetry (SDB) robustly across sites remains challenging. We analyze a Swin-Transformer based U-Net model (Swin-BathyUNet) to understand how it infers depth and when its predictions are trustworthy. A leave-one-band out study ranks spectral importance to the different bands consistent with shallow water optics. We adapt ablation-based CAM to regression (A-CAM-R) and validate the reliability via a performance retention test: keeping only the top-p% salient pixels while neutralizing the rest causes large, monotonic RMSE increase, indicating explanations localize on evidence the model relies on. Attention ablations show decoder conditioned cross attention on skips is an effective upgrade, improving robustness to glint/foam. Cross-region inference (train on one site, test on another) reveals depth-dependent degradation: MAE rises nearly linearly with depth, and bimodal depth distributions exacerbate mid/deep errors. Practical guidance follows: maintain wide receptive fields, preserve radiometric fidelity in green/blue channels, pre-filter bright high variance near shore, and pair light target site fine tuning with depth aware calibration to transfer across regions.
Paper Structure (27 sections, 15 equations, 3 figures, 5 tables)

This paper contains 27 sections, 15 equations, 3 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Motivation
  3. Challenges
  4. Contributions
  5. Related Work
  6. Dataset
  7. MagicBathyNet benchmark
  8. Experiments
  9. Experimental setup (Agia Napa).
  10. Data preprocessing and augmentation.
  11. Training/evaluation protocol.
  12. Attention Ablation Study for Swin-BathyUnet
  13. Motivation and Rationale
  14. Ablation Design Focused on Attention
  15. Supervision design: masks, losses, and training protocol.
  16. ...and 12 more sections

Figures (3)

  • Figure 1: Per-pixel views of Bands 0–2 and their RGB composite (2,1,0) for two example for Agia Napa region, illustrating shoreline-to-offshore gradients and band-wise noise/contrast differences.
  • Figure 2: Qualitative results on Agia Napa tiles. Left to right: (i) high-resolution aerial reference, (ii) co-registered S2 chip used by the network, (iii) Swin-BathyUNet bathymetry prediction (meters), (iv) LiDAR ground truth, and (v) A-CAMR saliency map. Top to bottom tile number 349, 359, 377, 398 for the Agia Napa region from MagicBathyNet dataset.
  • Figure 3: Depth-binned bathymetry error (lower is better). MAE is plotted against depth for the first 20 images of Puck Lagoon area in the MagicBathyNet dataset using the model trained on Agia Napa region; light-blue bars denote pixel counts on the secondary axis. Errors increase nearly linearly with depth, while a bimodal pixel distribution suggests data imbalance contributes to degradation at depth.