The MagicBathy project has delivered substantial scientific and technical progress toward enabling accurate, scalable, and multimodal shallow-water mapping through deep learning and remote sensing. The following key research activities and outcomes were achieved:
1. MagicBathyNet Dataset Development
A cornerstone contribution of the project is the creation of MagicBathyNet, the first publicly available multimodal benchmark dataset (available at
https://zenodo.org/records/10470959(si apre in una nuova finestra)) for joint bathymetry and seabed classification from optical imagery. The dataset consists of co-registered image patches from Sentinel-2, SPOT-6, and aerial platforms, along with bathymetry in raster format and corresponding pixel-level seabed class annotations. This enables supervised training and benchmarking of deep models under realistic multisensor scenarios.
In addition to labeled data, MagicBathyNet includes a large volume of unlabeled samples, facilitating self-supervised pretraining strategies prior to task-specific fine-tuning. This makes it a uniquely valuable resource for advancing representation learning in underwater and coastal remote sensing. The dataset has been curated across diverse coastal regions, capturing environmental variability and ensuring generalizability of trained models.
MagicBathyNet was exploited to benchmark state-of-the-art methods in learning-based bathymetry and pixel-based classification.
2. Sea-Undistort Synthetic Dataset
To overcome the lack of real-world paired data for evaluating and correcting through-water image distortion, we developed Sea-Undistort (available at
https://zenodo.org/records/15639838(si apre in una nuova finestra)) a synthetic dataset comprising 1200 high-resolution (512×512) image pairs rendered in Blender. Each pair includes a distortion-free and a distorted underwater view, with realistic effects such as sun glint, water surface waves, and scattering over a variety of seabed types.
Crucially, each image is accompanied by metadata such as camera parameters, sun position, and average depth, enabling highly controlled, supervised training of distortion-correcting models, including both generative and non-generative architectures. Sea-Undistort is the first of its kind, paving the way for foundational research in underwater image correction and self-supervised learning in domains where clean references are not physically attainable.
We used Sea-Undistort to benchmark state-of-the-art image restoration methods alongside our proposed variant; an enhanced lightweight diffusion-based framework with an early-fusion sun-glint mask. When applied to real aerial data, the proposed variant delivers more complete Digital Surface Models (DSMs) of the seabed, especially in deeper areas, reduces bathymetric errors, suppresses glint and scattering, and crisply restores fine seabed details.
3. Swin-BathyUNet architecture for Spectrally Derived Bathymetry
To address limitations in both Spectrally Derived Bathymetry (SDB) and SfM-MVS approaches - such as missing depth data in texture-poor regions - we proposed Swin-BathyUNet, a novel architecture that fuses U-Net with Swin Transformer-based self-attention and cross-attention mechanisms.
It can:
• Operate as a standalone SDB method, or
• Leverage incomplete SfM-MVS DSMs to learn complete bathymetric predictions.
The model captures long-range spatial dependencies and integrates spectral and geometric cues. Experiments in diverse test sites (Mediterranean and Baltic Seas) demonstrate that Swin-BathyUNet significantly improves bathymetric accuracy, completeness, and noise robustness over state-of-the-art methods, producing more reliable Digital Surface Models (DSMs) in challenging shallow-water environments.
4. Seabed-Net architecture: Joint Bathymetry and Seabed Classification
Recognizing the limitations of treating depth estimation and seabed classification as separate tasks, we introduced Seabed-Net, a unified multitask architecture that predicts both bathymetry and seabed class maps from multispectral remote sensing imagery.
Key features include:
• Dual-task encoder-decoder branches,
• An Attention Feature Fusion (AFF) module,
• A windowed Swin Transformer for multi-scale feature integration, and
• Dynamic task uncertainty weighting to balance learning across tasks.
Seabed-Net consistently outperforms single-task and multitask baselines, achieving 10–30% lower RMSE in depth estimation and up to 8% higher accuracy in seabed classification. The model enhances spatial coherence, improves detection in low-contrast substrates, and strengthens semantic boundary localization. These results confirm that joint modeling of bathymetry and substrate improves both tasks and enables richer environmental characterization.
5. Large-scale pretraining for Ocean Remote Sensing
The project explored the development of foundational models for ocean remote sensing by critically evaluating various self-supervised learning (SSL) paradigms, particularly:
• Contrastive vs. generative SSL,
• The integration of geolocation and water quality metadata, and
• Transfer performance on tasks like bathymetry estimation and marine debris detection.
This research contributes to a better understanding of how large-scale, unlabeled remote sensing datasets can be harnessed to pretrain general-purpose models capable of efficient fine-tuning on downstream ocean monitoring tasks.
6. Super-Resolution for Optical Imagery in Coastal Mapping
Lastly, research was conducted into super-resolution methods for enhancing the spatial detail of satellite and aerial imagery used in seabed mapping.
Both bathymetry-guided and agnostic variants were explored to evaluate their capacity to:
• Recover fine-scale seabed structure,
• Improve feature separation in mixed substrates, and
• Enable high-resolution predictions from low-resolution inputs (e.g. Sentinel-2).
Preliminary results show promising gains in restoring edge sharpness and improving model performance in downstream bathymetry and classification tasks.
