SKeMa: Satellite-based Kelp Mapping

Model Description

SKeMa is a Python-based tool for semantic segmentation of kelp in Sentinel-2 satellite imagery, developed for coastal monitoring. It uses a deep learning model from the Segmentation Models PyTorch library, with a MaxViT-Tiny-512 encoder and UNet decoder. The model outputs a binary GeoTIFF (kelp = 1, non-kelp = 0).

Authors: Mohsen Ghanbari, Neil Ernst, Taylor A. Denouden, Luba Y. Reshitnyk, Piper Steffen, Alena Wachmann, Eduardo Loos, Margot Hessing-Lewis, Maycira Costa
License: Creative Commons Attribution 4.0 International (CC-BY-4.0)
Citation: SKeMa models by Mohsen Ghanbari et al., SPECTRAL Remote Sensing Laboratory, University of Victoria, DOI: 10.57967/hf/6790
GitHub: m5ghanba/skema
PyPI: skema-kelp

Intended Use

Primary Use: Mapping kelp in Sentinel-2 imagery for environmental research and industry applications.
Users: Researchers, coastal ecologists, and industries analyzing Sentinel-2 data.
Out of Scope: Non-Sentinel-2 imagery or non-kelp segmentation tasks.

Model Architecture

Framework: PyTorch, Segmentation Models PyTorch (SMP) library.
Encoder: MaxViT-Tiny-512.
Decoder: UNet.
Input Channels: 12 (Sentinel-2 bands B2, B3, B4, B8, B5; substrate; bathymetry; spectral indices with the order NDVI; NDWI; GNDVI; ClI; NDVIRE).
Model Types:
- model_full: Uses all 12 channels, including bathymetry and substrate (optimized for BC coast).
- model_s2bandsandindices_only: Uses Sentinel-2 bands and indices only (no bathymetry/substrate).

Input Data

Required: Sentinel-2 .SAFE folder (download from Copernicus Browser).
For model_full: Bathymetry (Bathymetry_10m.tif) and substrate TIFFs (NCC_substrate_20m.tif, etc.) from BC coast sources (see GitHub README).

Output Format

Binary GeoTIFF (output.tif): Kelp = 1, non-kelp = 0.
Intermediate files:
- <SAFE_name>_B2B3B4B8.tif: 10m resolution, 4-band GeoTIFF (B02, B03, B04, B08).
- <SAFE_name>_B5B6B7B8A_B11B12.tif: 20m resolution, 6-band GeoTIFF.
- For model_full: <SAFE_name>_Bathymetry.tif, <SAFE_name>_Substrate.tif.

Usage Instructions

Installation via pip (Recommended)

pip install skema-kelp

Installation from source

Install:

git clone https://github.com/m5ghanba/skema.git
cd skema
python -m venv skema_env
source skema_env/bin/activate  # On Windows: skema_env\Scripts\activate
pip install .

See GitHub README for detailed setup.

Run:

skema --input-dir path/to/sentinel2/safe/folder --output-filename output.tif --model-type model_full
# Or for S2-only model:
skema --input-dir path/to/sentinel2/safe/folder --output-filename output.tif --model-type model_s2bandsandindices_only

# For help:
skema --help

Dependencies (from requirements.txt):
- PyTorch (GPU: torch==2.1.0, see GitHub for CUDA setup).
- Others: numpy, rasterio, segmentation-models-pytorch.

Training Data

Source: Sentinel-2 imagery from Copernicus Browser, BC coast bathymetry/substrate data.
Region: Optimized for British Columbia coastal waters.
Note: Bathymetry/substrate files required for model_full are available online (see GitHub README).

Limitations

Optimized for BC coast; performance may vary elsewhere.
model_full requires specific bathymetry/substrate data.
Assumes Sentinel-2 .SAFE format input.

Citation

Please cite as:

SKeMa model by Mohsen Ghanbari et al., SPECTRAL Remote Sensing Laboratory, University of Victoria, DOI: 10.57967/hf/6790.

BibTeX:

@software{skema_2025,
  author       = {Mohsen Ghanbari, Neil Ernst, Taylor A. Denouden, Luba Y. Reshitnyk, Piper Steffen, Alena Wachmann, Eduardo Loos, Margot Hessing-Lewis, Maycira Costa},
  title        = {SKeMa: Satellite-based Kelp Mapping using Semantic Segmentation on Sentinel-2 imagery},
  year         = 2025,
  publisher    = {Hugging Face},
  doi          = {10.57967/hf/6790},
  url          = {https://huggingface.co/m5ghanba/SKeMa}
}

m5ghanba
/

SKeMa