kelp-skema

Binary kelp canopy segmentation models for Sentinel-2 L2A imagery, developed by Mohsen Ghanbari (@m5ghanba) at the Spectral Lab, University of Victoria.. This repository hosts ONNX exports of the original SKeMa models (HuggingFace · GitHub), re-packaged for use with habitat-mapper. Normalization and spectral index computation are baked into the inference graph, so models accept raw Sentinel-2 DN values directly.

Model Variants

Four variants are available, differing in input channels and auxiliary data requirements:

Variant	ONNX file	Channels in	Auxiliary data	Geographic scope
S2-only	`Unet_tu-maxvit_tiny_tf_512_20260414.onnx`	5	None	Global
Full	`Unet_tu-maxvit_tiny_tf_512_20260414_full.onnx`	8	RF substrate, bathymetry, slope	Coastal BC
Ensemble	`Unet_tu-maxvit_tiny_tf_512_20260414_ensemble.onnx`	8	RF substrate, bathymetry, slope	Coastal BC
Full-BoPs	`Unet_tu-maxvit_tiny_tf_512_20260414_full_bops.onnx`	8	BoPs substrate, bathymetry, slope	Coastal BC

The Full and Full-BoPs variants use the same UNet weights but were trained with different substrate datasets. Use Full-BoPs for scenes acquired by BC Ministry of Environment BoPs sensors (identifiable by UXQ, UXS, or UDU in the Sentinel-2 product identifier). Use Full for all other scenes.

The Ensemble model combines the S2-only and Full models by averaging their logits, and generally produces the most robust predictions.

Architecture

All variants use a UNet with a MaxViT-tiny encoder (tu-maxvit_tiny_tf_512, tile size 512), implemented via segmentation-models-pytorch.

Each model's forward pass:

Accepts raw Sentinel-2 DN values (uint16 range, 0–~10000) as float32
Computes spectral indices internally (NDVI, NDWI, GNDVI, CIG, NDVIre)
Normalizes the augmented channel stack with baked-in per-channel mean/std
Runs the UNet backbone
Returns raw logits (apply sigmoid + threshold externally)

Internal channel layout after augmentation:

ch	S2-only (10 ch)	Full / Full-BoPs / Ensemble (13 ch)
0	B02 (Blue)	B02 (Blue)
1	B03 (Green)	B03 (Green)
2	B04 (Red)	B04 (Red)
3	B08 (NIR)	B08 (NIR)
4	B05 (Red Edge)	B05 (Red Edge)
5	NDVI	Substrate
6	NDWI	Bathymetry
7	GNDVI	Slope
8	CIG	NDVI
9	NDVIre	NDWI
10	—	GNDVI
11	—	CIG
12	—	NDVIre

Baked-in Normalization Parameters

Applied as (x - mean) / std after index augmentation.

S2-only (10 channels)

mean = [2.08900522e+02, 2.70272557e+02, 1.52312422e+02, 9.87182507e+02, 3.26650321e+02,
        5.65647882e-02, 1.62913280e-01, -1.62913280e-01, 1.63743045e+00, 1.09887867e-01]
std  = [1.59021908e+02, 2.18393833e+02, 2.08355086e+02, 1.29667310e+03, 3.79794112e+02,
        6.53314129e-01, 7.11820223e-01,  7.11820223e-01, 3.84363017e+00, 3.90619966e-01]

Full / Ensemble — RF substrate (13 channels)

mean = [2.02127847e+02, 2.64991799e+02, 1.45913497e+02, 9.57456953e+02, 3.20302883e+02,
        1.37548690e+00, -6.30723576e+00, 7.60650406e+00,
        3.66107438e-02, 1.84492036e-01, -1.84492036e-01, 1.56750584e+00, 9.99078992e-02]
std  = [1.61504107e+02, 2.22303637e+02, 2.03997451e+02, 1.26105656e+03, 3.79069759e+02,
        1.36767732e+00,  2.35930351e+02, 1.14107889e+01,
        6.71879776e-01, 7.23202999e-01,  7.23202999e-01, 3.86945642e+00, 4.06695959e-01]

Full-BoPs — BoPs substrate (13 channels, ch 5 differs)

mean = [2.02127847e+02, 2.64991799e+02, 1.45913497e+02, 9.57456953e+02, 3.20302883e+02,
        8.15331190e-01, -6.30723576e+00, 7.60650406e+00,
        3.66107438e-02, 1.84492036e-01, -1.84492036e-01, 1.56750584e+00, 9.99078992e-02]
std  = [1.61504107e+02, 2.22303637e+02, 2.03997451e+02, 1.26105656e+03, 3.79069759e+02,
        1.09273473e+00,  2.35930351e+02, 1.14107889e+01,
        6.71879776e-01, 7.23202999e-01,  7.23202999e-01, 3.86945642e+00, 4.06695959e-01]

Repo Files

20260414/
├── Unet_tu-maxvit_tiny_tf_512_20260414.onnx          # S2-only model
├── Unet_tu-maxvit_tiny_tf_512_20260414_full.onnx     # Full model (RF substrate)
├── Unet_tu-maxvit_tiny_tf_512_20260414_ensemble.onnx # Ensemble model (RF substrate)
├── Unet_tu-maxvit_tiny_tf_512_20260414_full_bops.onnx # Full model (BoPs substrate)
├── substrate_20m_cog.tif                              # RF substrate raster (20 m)
├── bops_substrate_10m_cog.tif                         # BoPs substrate raster (10 m)
├── bathymetry_10m_cog.tif                             # Bathymetry raster (10 m)
└── slope_10m_cog.tif                                  # Slope raster (10 m)

Usage

The recommended way to run these models is via habitat-mapper:

hab predict <input.SAFE> <output.tif> --model kelp-skema

Model configs are bundled in habitat-mapper and handle dependency downloading, band reading, tiling, and postprocessing automatically. Four configs are available corresponding to the four variants:

kelp-skema/20260414 — S2-only
kelp-skema/20260414-full — Full (RF substrate)
kelp-skema/20260414-ensemble — Ensemble
kelp-skema/20260414-full-bops — Full (BoPs substrate)

Direct ONNX inference

The models accept batched float32 input of shape (B, C, H, W) where C is 5 (S2-only) or 8 (full/ensemble/full-bops). Values should be raw Sentinel-2 L2A DN values with the processing baseline offset removed (subtract 1000 for baseline ≥ 4.0, leave as-is for older baselines, then clip to 0).

import onnxruntime as ort
import numpy as np

session = ort.InferenceSession("Unet_tu-maxvit_tiny_tf_512_20260414_full.onnx")
logits = session.run(["output"], {"input": tile.astype(np.float32)})[0]
predictions = (1 / (1 + np.exp(-logits)) > 0.5).astype(np.uint8)

Intended Use

These models are intended for mapping surface kelp canopy extent from multispectral Sentinel-2 satellite imagery, primarily in coastal British Columbia, Canada. They are not validated for other species or regions.

Limitations

The Full, Ensemble, and Full-BoPs variants require static auxiliary layers (substrate, bathymetry, slope) that are only available for coastal British Columbia.
All variants are trained on Sentinel-2 L2A (atmospherically corrected) data. Results on L1C data are not validated.
Performance degrades under heavy cloud cover, sun glint, or turbid water conditions typical of intertidal zones.
Canopy kelp only — subsurface kelp is not detectable from multispectral imagery.

Citation

If you use these models in your work, please cite the following papers:

@article{ghanbari2026ecoinf,
  title     = {Automated Kelp Forest Mapping from Sentinel-2 Imagery Using Deep Learning},
  author    = {Ghanbari, Mohsen and Ernst, Neil and Denouden, Taylor and Reshitnyk, Luba
               and Steffen, Piper and Wachmann, Alena and Mora-Soto, Alejandra
               and Loos, Eduardo and Hessing-Lewis, Margot and Dedeluk, Nic and Costa, Maycira},
  journal   = {Ecological Informatics},
  year      = {2026},
  note      = {Under review, Manuscript No. ECOINF-D-26-01969}
}

The original model weights and inference code are available at:

License

MIT — see LICENSE.

Downloads last month: -; Downloads are not tracked for this model. How to track