DL_bathy pretrained model weights
Temporary home: this repo currently lives under a personal account (
Amy5133) while the author arranges access to thebuckai-observatoryHugging Face org thatDL_bathy'sinference.pydefaults to. Expect this to move tobuckai-observatory/dl-bathy-weightslater;inference.py'sHF_DEFAULT_REPO_IDwill be updated to match when it does.
Pretrained weights for the five models compared in Hsu & Moortgat (2026),
From Local Training to Large-Scale Mapping: A Comparative Assessment of
Machine Learning and Deep Learning for Transferable Satellite-Derived
Bathymetry, Remote Sensing 18(11): 1768,
10.3390/rs18111768. Trained and
evaluated using the DL_bathy pipeline:
github.com/buckai-observatory/DL_bathy.
Repository layout
./
βββ best_model.pth ConvNeXt-Large, best-val checkpoint (default for inference.py)
βββ deeplabv3_bathy_final.pth ConvNeXt-Large, final-epoch checkpoint
βββ metadata.json ConvNeXt-Large training run config + metrics
βββ resnet101/
β βββ best_model.pth
β βββ deeplabv3_bathy_final.pth
β βββ metadata.json
βββ resnet50/
β βββ best_model.pth
β βββ deeplabv3_bathy_final.pth
β βββ metadata.json
βββ efficientnet_b4/
β βββ best_model.pth
β βββ deeplabv3_bathy_final.pth
β βββ metadata.json
βββ random_forest/
βββ trained_model_20260126_173237.joblib
For each DeepLabV3+ backbone, both the best-validation-metric checkpoint
(best_model.pth) and the final-epoch checkpoint
(deeplabv3_bathy_final.pth) are included, alongside the training run's
metadata.json (configuration + metrics, written by
step2_train_test.py). All four were trained with the SWF (Smooth
Weight Function) depth-weighted loss (loss_type: "swf" in each
metadata.json).
Load a .pth checkpoint with segmentation_models_pytorch via the
DL_bathy repo's inference.py (load_model(filename=..., encoder_name=...)); load the .joblib file with joblib.load() into
scikit-learn.
Model Card β DL_bathy (satellite-derived bathymetry)
Overview
- Model name: DL_bathy DeepLabV3+ satellite-derived bathymetry
- Model type: fully-convolutional segmentation network (regression head)
- Task: per-pixel water depth estimation (metres) from Sentinel-2 L2A imagery
- Architecture:
smp.DeepLabV3Pluswith a 12-band input head, single-channel regression output, and an ImageNet-pretrained encoder - Default encoder:
tu-convnext_large(ConvNeXt-Large, timm backend). Alternate encoders trained by the same pipeline:efficientnet-b4,resnet50,resnet101. - License: MIT (matches the DL_bathy repository)
Intended use
- Predicting shallow-water bathymetry (depth in metres) over any location covered by Sentinel-2, given a cloud-free (or SCL-masked) L2A tile containing bands B01βB12.
- Research and educational use: reproducing the results of Hsu & Moortgat (2026); benchmarking against other satellite-derived bathymetry (SDB) approaches; exploring transfer to new geographies.
Out-of-scope / cautions
- Deep water: the training distribution is shallow-water reef and coastal environments (< ~25 m). Predictions for open ocean or turbid inland water bodies are extrapolations and should not be trusted without independent validation.
- Turbidity, sun glint, breaking waves: not explicitly handled by the model; masking these upstream (e.g. using SCL or a scene-level QA step) is the caller's responsibility.
- Not a real-time / navigational product: outputs are research-grade, not certified for maritime navigation, dredging planning, or any safety-critical use.
Training data
- Ground truth: airborne LiDAR bathymetry (see Hsu & Moortgat, 2026 for the full site list and processing chain).
- Public reference subset: an open subset of the ground-truth data
is shipped in
Ground_Truth.zipin the DL_bathy repository for reproducibility and demo purposes. The full training dataset includes commercially licensed LiDAR that cannot be redistributed; this is documented in the repo'sData availabilitysection. - Sentinel-2 imagery: L2A reflectance bands 1β12 (B01βB12), fetched at the training AOIs / dates via the geoai-datacubes pipeline. Cloud and shadow pixels are excluded via the SCL band during training.
TODO: total number of training patches, patch size (px), split ratios, augmentation strategy, and per-site sample counts.
Training configuration
See each architecture's metadata.json in this repo for the exact
per-run configuration (loss type, SWF hyperparameters, learning rate,
epochs trained, best validation loss, etc.). Summary common to all four:
- Input channels: 12 (Sentinel-2 bands B01βB12)
- Output: single-channel regression, no activation (raw depth in metres)
- Loss: SWF (depth-weighted RMSE) β see
SWF_beta/SWF_Z0in eachmetadata.json - Optimiser: AdamW, weight-decay 1e-4
- Scheduler: ReduceLROnPlateau (factor 0.5)
- Hardware / wall-clock time: TODO
Evaluation
TODO β from the release training run:
- Held-out test set: RMSE (m), MAE (m), RΒ², bias
- Per-depth-bin performance (e.g. RMSE at 0β5 m, 5β10 m, 10β15 m, > 15 m)
- Transfer performance on a geographically-held-out site, if measured
How to use
Minimal example (needs pip install huggingface_hub geoai-datacubes and
the DL_bathy repository on the Python path):
from inference import predict_from_geoai_datacubes
# Predict bathymetry over a small AOI in the Great Barrier Reef (the same
# region the public Ground_Truth.zip covers).
depths, meta = predict_from_geoai_datacubes(
aoi=[145.14, -14.56, 145.66, -14.26],
time_range=("2020-06-01", "2020-09-30"),
max_cloud_coverage=0.20,
)
print(f"Predicted depth map shape: {depths.shape}, dtype: {depths.dtype}")
To load a non-default architecture:
from inference import load_model, predict_depth
model = load_model(filename="resnet101/best_model.pth", encoder_name="resnet101")
See notebooks/dl_bathy_demo.ipynb in the DL_bathy repository for a
Colab-ready end-to-end walkthrough on the public GBR subset.
Citation
If you use these weights, please cite:
- Hsu, H.-J., & Moortgat, J. (2026). From Local Training to Large-Scale Mapping: A Comparative Assessment of Machine Learning and Deep Learning for Transferable Satellite-Derived Bathymetry. Remote Sensing 18(11): 1768. https://doi.org/10.3390/rs18111768
- The DL_bathy software: see
CITATION.cffin the DL_bathy repository.
Version history
v0.1.0β first upload, reproducing the setup used in Hsu & Moortgat (2026).