skema-kelp 0.3.5

Satellite-based kelp classification from Sentinel-2 using semantic segmentation

SKeMa

Satellite-based Kelp Mapping using Semantic Segmentation on Sentinel-2 imagery

skema is a Python tool for classifying kelp in Sentinel-2 satellite images using a deep learning segmentation model (PyTorch). It provides a command-line interface (CLI) for easy, reproducible inference. To run the tool you would need to download Sentinel-2 images from the Copernicus Browser. More detailed instruction on how to download these images can be found in Section Usage. The following instructions are provided for anyone with no knowledge of what a command line is, no knowledge of Python or virtual environments, etc. Just follow along step by step.

Model available on Hugging Face: m5ghanba/SKeMa
DOI: 10.57967/hf/6790

---

Table of Contents

• Quick Start (Experienced Users)
• Citation
• Installation
  • Step 1: Install Python
  • Step 2: Install SKeMa
• Usage
  • Activating the Virtual Environment
  • Downloading Sentinel-2 Images
  • Running SKeMa
  • Model Types
  • Output Files
  • Opening and Visualizing Kelp Maps in QGIS
• Static Files
• GPU Support
• Project Structure
• License

---

Quick Start (Experienced Users)

pip install skema-kelp

# Download static files (for model_full only) from sources listed below
# Download Sentinel-2 imagery from https://dataspace.copernicus.eu/browser/

skema --input-dir "path/to/S2_scene.SAFE" --output-filename output.tif

# For help and all options
skema --help

For detailed installation instructions (beginner-friendly), see Installation below.

---

Citation

If you use SKeMa in your research or work, please cite:

@software{skema_2026,
  author       = {Mohsen Ghanbari, Neil Ernst, Taylor A. Denouden, Luba Y. Reshitnyk, Piper Steffen, Alena Wachmann, Alejandra Mora-Soto, Eduardo Loos, Margot Hessing-Lewis, Nic Dedeluke, Maycira Costa}, 
  title        = {SKeMa: Satellite-based Kelp Mapping using Semantic Segmentation on Sentinel-2 imagery},
  year         = 2026,
  publisher    = {Hugging Face},
  doi          = {10.57967/hf/6790},
  url          = {https://huggingface.co/m5ghanba/SKeMa}
}

---

Installation

Before you can set up SKeMa, you'll need Python (version 3.9 to 3.12) installed on your computer. Python is a free tool, and no accounts or sign-ups are required to install it. We'll install it using your terminal (command line) where possible for simplicity. If you're on Windows, ensure you're using PowerShell or Command Prompt as Administrator (right-click and select "Run as administrator") for some steps. In some cases, however, you may need to use Command Prompt instead of PowerShell because PowerShell can handle command resolution and PATH variables differently, which may prevent it from recognizing Python even if it is installed. Additionally, if Python was installed only for your user account (and not system-wide), it may only be accessible from a normal Command Prompt session rather than one opened as Administrator, since elevated terminals can use a different set of environment variables.

Step 1: Install Python

Checking Your Current Python Version:

Before proceeding, check if you already have Python between versions 3.9 and 3.12 on your computer. Open a terminal (PowerShell, Command Prompt, or macOS/Linux terminal).

Run:

python --version

or, on macOS/Linux, try:

python3 --version

If the version shown is between 3.9 and 3.12 (e.g., Python 3.9.13, Python 3.11.5, Python 3.12.7), you already meet the requirement and can skip the installation steps below. Otherwise, follow the instructions below to install Python.

On Windows

1. Download the Python 3.12.7 installer directly from the Python website:

   • https://www.python.org/ftp/python/3.12.7/python-3.12.7-amd64.exe

2. Run the installer:

   • IMPORTANT: At the very first screen, check the box that says "Add Python to PATH". This option is unchecked by default — if you miss it, Python will not be recognized by your terminal.
   • Click "Install Now" and follow the prompts.

3. Verify the installation:

   • Open Command Prompt (search for cmd in the Start menu). Use Command Prompt, not PowerShell, as PowerShell can sometimes fail to recognise Python even if it is installed.
   • Run python --version. It should output Python 3.12.7.
   • If you see an error, close and reopen Command Prompt and try again. If it still fails, search online for "add Python to PATH Windows".

On macOS

1. Install Homebrew (a package manager for CLI installations, if you don't have it):

   • Open Terminal (search for it in Spotlight with Cmd+Space).
   • Run:

     /bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/HEAD/install.sh)"
     

   • Follow any on-screen prompts (it may ask for your password; this is normal). No account needed.
   • After installation, run the commands it suggests to add Homebrew to your PATH (e.g., echo 'eval "$(/opt/homebrew/bin/brew shellenv)"' >> ~/.zprofile and then eval "$(/opt/homebrew/bin/brew shellenv)").
   • Verify: Run brew --version. It should show a version like "4.3.0".

2. Install Python 3.12:

   • Run:

     brew install python@3.12
     

   • This installs Python and adds it to PATH.
   • Verify: Run python3 --version (note: use python3 on macOS). It should output "Python 3.12.7".

   Alternative: Download the official installer from python.org using your browser, run it, and follow GUI steps. Make sure Python is added to your PATH during installation.

On Linux (e.g., Ubuntu/Debian; adjust for other distros like Fedora)

1. Update your package list:

   • Open your terminal.
   • Run:

     sudo apt update
     

   • Enter your password when prompted (sudo is for admin privileges; no account needed beyond your user login).

2. Install Python 3.12:

   • Run:

     sudo apt install python3.12 python3.12-venv python3-pip
     

   • This installs Python, the venv module, and pip.
   • Verify: Run python3 --version. It should output "Python 3.12.x".
   • For Fedora/RHEL: Use sudo dnf install python3.12 instead.

   Note: If your distro's repositories don't have Python 3.12, add a PPA (e.g., for Ubuntu: sudo add-apt-repository ppa:deadsnakes/ppa then update and install).

Once Python is installed and verified, proceed to the next section. If you encounter errors (e.g., "command not found"), search online for the exact error message + your OS.

Step 2: Install SKeMa

Open your terminal:

• On Windows, you can use Command Prompt or PowerShell.
• On macOS, open the Terminal app.
• On Linux, open your terminal emulator of choice.

When you open a terminal, you start inside a directory (folder). You can move to another directory with the command cd. For example:

cd C:\Users\YourName\Documents  

On macOS/Linux:

cd /Users/yourname/Documents  

👉 The easiest way to navigate is to open your file explorer, go to the folder you want, then copy its full path and paste it after cd on the command line. For more details, look up "basic terminal navigation" online.

Now, navigate to a directory where you want to work with SKeMa, then run:

Option 1: Install with pip (Recommended)

# Create a virtual environment 
python -m venv skema_env

# Activate the virtual environment
# On Windows:
skema_env\Scripts\activate
# On macOS/Linux:
source skema_env/bin/activate

# Install SKeMa
pip install skema-kelp

⚠️ Windows only — "Access Denied" when activating the environment

This is caused by PowerShell's execution policy, which blocks scripts from running by default on many Windows systems. Fix it by running this command once in PowerShell:

Set-ExecutionPolicy -Scope CurrentUser -ExecutionPolicy RemoteSigned

Type Y and press Enter when prompted. This only affects your own user account and does not change system-wide security settings. After running it, try activating the environment again.

⏳ Installation Time: The installation may take up to 10 minutes to complete. A large amount of text will scroll past — this is normal. You will know it is finished when you see a line like Successfully installed skema-kelp-X.X.X and your command prompt reappears.

⚠️ Opencv issue for macOS users: In some cases, the installation may fail while building the opencv-python-headless dependency (a computer vision library required by albumentations). If you see an error mentioning OpenCV or "Failed building wheel for opencv-python-headless", install OpenCV separately first and then install SKeMa again:

pip install opencv-python-headless==4.9.0.80
pip install skema-kelp

This installs a precompiled OpenCV wheel for macOS and avoids a lengthy build process.

Option 2: Install from source (for developers)

If you want to modify the code or contribute to development:

# Install Git first (see system-specific instructions below)
# Then clone the repository
git clone https://github.com/m5ghanba/skema.git
cd skema

# Create and activate virtual environment
python -m venv skema_env
# On Windows:
skema_env\Scripts\activate
# On macOS/Linux:
source skema_env/bin/activate

# Install in development mode
pip install -e .

Alternatively, users can use the provided poetry.lock file with Poetry (poetry install), which will automatically create and manage a virtual environment while ensuring all dependencies are installed with exact, reproducible versions.

Installing Git (only needed for Option 2):

• Windows: winget install --id Git.Git -e --source winget or download from git-scm.com
• macOS: brew install git or download from git-scm.com
• Linux: sudo apt install git (Ubuntu/Debian) or sudo dnf install git (Fedora/RHEL)

Each line explained:

• python -m venv skema_env: Creates a virtual environment named skema_env to isolate project dependencies.
• skema_env\Scripts\activate (Windows) or source skema_env/bin/activate (macOS/Linux): Activates the virtual environment, ensuring subsequent commands use its isolated Python and packages.
• pip install skema-kelp: Installs SKeMa and all its dependencies from PyPI.

If you encounter packaging errors, make sure your pip and build tools are up to date:

pip install --upgrade pip setuptools wheel

---

Usage

Activating the Virtual Environment

To use SKeMa after the initial installation, you must activate its virtual environment each time you start a new session (if you created one). Follow these steps each time you want to run the tool:

1. Open a terminal (Command Prompt, PowerShell, or Terminal).
2. Navigate to the directory where you created the virtual environment:
   • On Windows:

     cd path\to\your\directory
     

   • On macOS/Linux:

     cd path/to/your/directory
     

3. Activate the virtual environment:
   • On Windows:

     skema_env\Scripts\activate
     

   • On macOS/Linux:

     source skema_env/bin/activate
     

4. Run SKeMa using the commands described below.

If your command line prompt shows (skema_env), the virtual environment is activated and you're ready to proceed.

Downloading Sentinel-2 Images

SKeMa uses Sentinel-2 satellite images, which can be downloaded from the Copernicus Browser. You will need a free account to access and download these images, which are provided as .zip files.

Sentinel-2 Image Download Instructions

Follow these steps to download Sentinel-2 images:

1. Go to the Copernicus Browser: Navigate to https://browser.dataspace.copernicus.eu/

2. Register: If you haven't already, create a free account by clicking on the "Register" option.

3. Sign in: Log in to your account using your credentials.

4. Create an area of interest: Click on the top right option "Create an area of interest" (the pentagon shape icon).

   

5. Draw your polygon: Click on the option "Draw polygon of interest..." (the pencil shape icon).

   

6. Define your area: Draw a polygon around your area of interest by clicking on the map to create points. Close the polygon by clicking on the first point you placed.

   

7. Set parameters and search:

   • Select the Time Range option and set the dates in the "From" and "Until" fields.
   • Optionally, set the cloud coverage filter.
   • Click on "Find products within selected time range".
   • Verify that Sentinel-2 L2A is selected as the data level (default setting).

   

8. Choose your Sentinel-2 scene: From the map, select the desired Sentinel-2 scene by clicking on it.

   

9. Preview and download:

   • Click on the "i" option to view a preview of the image and quick information.
   • Click on the download icon (located at the bottom right) to download the scene.

   

Running SKeMa

SKeMa can be run on a single Sentinel-2 scene or on a directory of multiple scenes at once.

Single Scene

skema --input-dir "path/to/sentinel2/scene.SAFE" --output-filename output.tif  

• --input-dir must be the full path to the .SAFE folder.

  • Sentinel-2 images from the Copernicus Browser come as .zip files. Extract them first.
  • Then, pass the full path to the .SAFE folder (e.g., "C:\...\S2C_MSIL2A_20250715T194921_N0511_R085_T09UUU_20250716T001356.SAFE").
  • Note: If your path or output filename contains spaces, enclose it in double quotation marks (as shown in the example).

• --output-filename is the name of the output file (e.g., output.tif). You only need to provide the filename, not the full directory path. The output will be saved in a folder created alongside the .SAFE folder.

Batch Mode (Multiple Scenes + Mosaic)

If you have multiple Sentinel-2 scenes to process, use the --batch-dir flag instead of running the tool repeatedly. Place all your unzipped .SAFE folders inside a single directory and point --input-dir to that directory:

skema --input-dir "path/to/folder/with/safe/files/" --output-filename output.tif --batch-dir

SKeMa will:

1. Detect all .SAFE folders inside the directory automatically.
2. Process each scene individually, exactly as in single-scene mode.
3. Save each scene's prediction inside a subfolder named after that scene. The output filename will be the scene name with your --output-filename value appended as a suffix. For example, if the scene is S2B_MSIL2A_20220806T191919_N0510_R099_T10UCV_20240718T204439.SAFE and --output-filename is output.tif, the prediction will be saved as:

   path/to/folder/with/safe/files/
   └── S2B_MSIL2A_20220806T191919_N0510_R099_T10UCV_20240718T204439/
       ├── S2B_MSIL2A_20220806T191919_N0510_R099_T10UCV_20240718T204439_B2B3B4B8.tif
       ├── S2B_MSIL2A_20220806T191919_N0510_R099_T10UCV_20240718T204439_B5B6B7B8A_B11B12.tif
       ├── ...
       └── S2B_MSIL2A_20220806T191919_N0510_R099_T10UCV_20240718T204439_output.tif
   

4. After all scenes are processed, generate a single mosaic by merging all predictions into one GeoTIFF saved as mosaic_kelp_map.tif in the input directory:

   path/to/folder/with/safe/files/
   ├── S2B_MSIL2A_.../
   ├── S2B_MSIL2A_.../
   └── mosaic_kelp_map.tif   ← combined prediction for all scenes
   

The mosaic is reprojected to BC Albers (EPSG:3005) at 10 m resolution (this is important if skema is applied to other regions in the world). Overlapping pixels between scenes are resolved by taking the maximum value, meaning a pixel is classified as kelp if any contributing scene predicted kelp there.

Model Types

SKeMa supports a few model types, selectable via the --model-type flag:

1. model_s2bandsandindices_only (default): Uses only Sentinel-2 bands and derived spectral indices. This model does not require bathymetry or substrate files, making it suitable for areas outside British Columbia or when these static files are unavailable.

2. model_full: Uses Sentinel-2 bands, as well as bathymetry, and substrate information.

3. model_ensemble: An ensemble model that combines predictions from both model_full and model_s2bandsandindices_only by averaging their outputs. This model requires the same static files as model_full (bathymetry and substrate) and can provide more robust predictions by leveraging both modeling approaches.

If --model-type is not specified, the tool defaults to model_s2bandsandindices_only.

⚠️ Static Files

As noted above, model_full and model_ensemble require additional static files. These files must be downloaded and placed in the appropriate folder within the SKeMa installation directory. Detailed instructions are provided in Static Files Section. In practice, the default model (model_s2bandsandindices_only) achieves accuracy that is often comparable to model_full and model_ensemble. Therefore, it is recommended when minimizing computation time and resource usage is important.

Optional Flags

--use-bops-substrate: By default, model_full and model_ensemble use a substrate layer derived from a Random Forest (RF) model. When --use-bops-substrate is set, SKeMa instead uses substrate layers from the Bottom Patches (BoPs) dataset. Each substrate source has its own trained model weights, which are downloaded automatically.

--soft-substrate-masking: When set, kelp pixels that overlap with sandy or muddy substrate classes are reclassified to 0 (no kelp). Use this with care - depending on substrate data quality in your area, it may remove a notable number of true kelp pixels significantly increasing omission error while providing only a modest reduction in commission error.

--eelgrass-masking: When set, kelp pixels that fall within eelgrass polygons from the BC Marine Conservation Analysis (BCMCA) eelgrass dataset are reclassified to 0 (no kelp). Eelgrass (Zostera marina) is a marine vascular plant that provides important habitat. Kelp predictions overlapping with it are likely false positives. This flag is specific to British Columbia. It can be combined with --soft-substrate-masking.

When either or both masking flags are set, a single combined _masked.tif output is produced (rather than separate files per flag). In batch mode, a mosaic_kelp_map_masked.tif is also created.

Usage Examples

Note: Scroll right to see the complete command if it extends beyond your screen.

# Default: S2 bands and indices only (no static files required)
skema --input-dir "path/to/sentinel2/safe/folder" --output-filename output.tif

# Full model with RF substrate (default substrate source)
skema --input-dir "path/to/sentinel2/safe/folder" --output-filename output.tif --model-type model_full

# Full model with BoPs substrate
skema --input-dir "path/to/sentinel2/safe/folder" --output-filename output.tif --model-type model_full --use-bops-substrate

# Ensemble model with BoPs substrate
skema --input-dir "path/to/sentinel2/safe/folder" --output-filename output.tif --model-type model_ensemble --use-bops-substrate

# Soft substrate masking (produces a combined _masked.tif output - requires static files)
skema --input-dir "path/to/sentinel2/safe/folder" --output-filename output.tif --soft-substrate-masking

# Apply eelgrass masking (works with any model type)
skema --input-dir "path/to/sentinel2/safe/folder" --output-filename output.tif --eelgrass-masking

# Combine both masking filters (single combined _masked.tif output - requires static files)
skema --input-dir "path/to/sentinel2/safe/folder" --output-filename output.tif --soft-substrate-masking --eelgrass-masking

# Batch mode with ensemble model and BoPs substrate
skema --input-dir "path/to/folder/with/safe/files/" --output-filename output.tif --batch-dir --model-type model_ensemble --use-bops-substrate

Output Files

Single Scene

After running, the tool generates a folder with the same name as the .SAFE file (without the .SAFE extension), located alongside it. Inside this folder, you'll find:

For model_s2bandsandindices_only:

1. <SAFE_name>_B2B3B4B8.tif: a 10 m resolution, 4-band GeoTIFF containing Sentinel-2 bands B02 (Blue), B03 (Green), B04 (Red), and B08 (Near-Infrared).
2. <SAFE_name>_B5B6B7B8A_B11B12.tif: a 20 m resolution, 6-band GeoTIFF containing Sentinel-2 bands B05, B06, B07, B8A, B11, and B12.
3. output.tif (or the filename you specify): a binary GeoTIFF, where kelp is labeled as 1 and non-kelp as 0.

For model_fulland model_ensemble:

1. <SAFE_name>_B2B3B4B8.tif: a 10 m resolution, 4-band GeoTIFF containing Sentinel-2 bands B02 (Blue), B03 (Green), B04 (Red), and B08 (Near-Infrared).
2. <SAFE_name>_B5B6B7B8A_B11B12.tif: a 20 m resolution, 6-band GeoTIFF containing Sentinel-2 bands B05, B06, B07, B8A, B11, and B12.
3. <SAFE_name>_Bathymetry.tif: bathymetry data aligned and warped to the Sentinel-2 pixel grid.
4. <SAFE_name>_Substrate.tif: substrate classification data aligned and warped to the Sentinel-2 pixel grid.
5. <SAFE_name>_Slope.tif: slope data (derived from bathymetry) aligned and warped to the Sentinel-2 pixel grid.
6. output.tif (or the filename you specify): a binary GeoTIFF, where kelp is labeled as 1 and non-kelp as 0.
7. output_masked.tif (only when --soft-substrate-masking and/or --eelgrass-masking are set): a single combined binary GeoTIFF with all selected masks applied. Kelp pixels on sandy/muddy substrate and/or within eelgrass polygons are set to 0.

Batch Mode

In addition to the per-scene output folders described above (with <SAFE_name>_output.tif inside each), batch mode produces one additional file in the input directory:

• mosaic_kelp_map.tif: a single binary GeoTIFF mosaic merging all scene predictions, in BC Albers projection (EPSG:3005) at 10 m resolution.

• mosaic_kelp_map_masked.tif (only when --soft-substrate-masking and/or --eelgrass-masking are set): a combined masked mosaic built from the per-scene _masked.tif files, in BC Albers projection (EPSG:3005) at 10 m resolution.

Opening and Visualizing Kelp Maps in QGIS

The output GeoTIFF files generated by SKeMa can be viewed and analyzed in GIS software such as QGIS (free and open source: https://qgis.org/). This section explains how to visualize the Sentinel-2 imagery in false color and overlay the predicted kelp map.

Opening the Sentinel-2 Image

The file:

<SAFE_name>_B2B3B4B8.tif

contains four Sentinel-2 bands at 10 m resolution:

Band	Description
B02	Blue
B03	Green
B04	Red
B08	Near Infrared (NIR)

To display this image as a false-color composite in QGIS:

1. Open QGIS.
2. Import: Drag and drop the file <SAFE_name>_B2B3B4B8.tif into the QGIS window.
3. Layer Properties: In the Layers panel, right-click the layer and select Properties.
4. Set Symbology: Navigate to the Symbology tab and configure the following:
   • Render type: Multiband color
   • Red band: Band 4 (B08 / NIR)
   • Green band: Band 3 (B04 / Red)
   • Blue band: Band 2 (B03 / Green)

This produces a standard Sentinel-2 false-color image where vegetation and kelp appear in shades of red.

Improving the Visualization with Contrast Stretching

Sentinel-2 reflectance values often appear dark by default. To improve visualization:

1. In the Symbology window, locate the Min / Max Value Settings.
2. Set Contrast enhancement to Stretch to MinMax.
3. Manually set the approximate maximum values for better clarity:
   • Red (NIR / B08): 400
   • Green (Red / B04): 600
   • Blue (Green / B03): 800
4. Click Apply.

These values work well for many coastal Sentinel-2 scenes, although optimal values may vary slightly between images depending on atmospheric conditions, season, and illumination.

Opening the Kelp Prediction Map

The prediction file (e.g., output.tif) is a binary raster:

Value	Meaning
1	Kelp
0	Non-kelp

To visualize the kelp overlay:

1. Import: Drag the prediction GeoTIFF into QGIS.
2. Layer Properties: Right-click the prediction raster and select Properties > Symbology.
3. Configure Overlay:
   • Set Render type to Paletted/Unique values.
   • Click Classify.
   • Value 1 (Kelp): Set fill color to yellow.
   • Value 0 (Non-kelp): Set opacity to 0% (transparent) after double-clicking on the color.
4. Click Apply.

The kelp canopy predictions will now appear as yellow regions overlaid on the Sentinel-2 false-color image.

For best visualization:

• Place the kelp layer above the Sentinel-2 image in the Layers panel.
• Optionally reduce kelp layer opacity slightly (e.g., 70–80%) to better see the underlying imagery.

---

🗂️ Static Files

These files are only required when using model_full or model_ensemble. If you are using the default model_s2bandsandindices_only, you can skip this section entirely.

Static files are bathymetry and substrate rasters for the British Columbia coast that SKeMa uses as additional input channels when predicting kelp.

Bathymetry and Slope Files

• Bathymetry.tif — a single TIFF raster of seafloor depth.
• Slope.tif — a single TIFF raster derived from the bathymetry data.

Substrate Files

SKeMa supports two substrate sources, selected via --use-bops-substrate:

1. Regional RF substrate files (20 m resolution) — default:

• Five TIFF rasters: NCC_substrate_20m.tif, SOG_substrate_20m.tif, WCVI_substrate_20m.tif, QCS_substrate_20m.tif, HG_substrate_20m.tif
• Each covers a different region of the BC coast.

2. BoPs substrate files (10 m resolution) — with --use-bops-substrate:

• Four TIFF rasters: BoPs_HG_10m.tif, BoPs_NCC_10m.tif, BoPs_QCSSOG_10m.tif, BoPs_WCVI_10m.tif
• Must be rasterized from shapefiles using the provided notebook notebooks/rasterizeNearshoreBottomPatches_BoPs.ipynb.

Where to place static files

Place all files in the bathy_substrate folder inside the SKeMa package:

Windows (pip install):

skema_env\Lib\site-packages\skema\static\bathy_substrate

macOS/Linux (pip install):

skema_env/lib/python3.12/site-packages/skema/static/bathy_substrate/

Cloned from GitHub:

skema/static/bathy_substrate/

If you encounter any issues downloading or placing these files, please contact us for assistance.

Sources

• Bathymetry — Canadian Hydrographic Service (CHS), 20 m resolution:

  • Documentation: https://publications.gc.ca/collections/collection_2019/mpo-dfo/Fs97-6-3321-eng.pdf
  • Dataset: https://www.gis-hub.ca/dataset/bathy_20m

• RF Substrate — Shallow substrate model (20 m) of the Pacific Canadian coast (Haggarty et al., 2020): https://osdp-psdo.canada.ca/dp/en/search/metadata/NRCAN-FGP-1-b100cf6c-7818-4748-9960-9eab2aa6a7a0

• BoPs Substrate — Nearshore Bottom Patches:

  • Full dataset: https://open.canada.ca/data/en/dataset/6cda0f8d-110e-423d-8d7a-bf8a40eaa26e
  • Report (English): https://publications.gc.ca/collections/collection_2022/mpo-dfo/Fs97-6-3472-eng.pdf

---

🖥️ GPU Support

Most users will not need this section. SKeMa runs on CPU by default and works well without a GPU. GPU support is optional and only relevant if you have an NVIDIA GPU and want faster inference.

Check your CUDA version:

nvcc --version

Then install the matching PyTorch build:

For CUDA 12.1:

pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121

For CUDA 11.8:

pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118

---

⚙️ Project Structure

skema/                                  ← repo root
│
├── .github/
│   └── workflows/
│       └── tests.yml                   ← CI: runs pytest on every push / PR
│
├── skema/                              ← installable Python package
│   ├── __init__.py                     ← exposes segment()
│   ├── cli.py                          ← Click entry-point (skema command)
│   ├── lib.py                          ← high-level orchestration (segment, create_mosaic)
│   │
│   ├── preprocessing/                  ← all input-preparation logic
│   │   ├── __init__.py
│   │   ├── band_extraction.py          ← extract S2 bands from .SAFE → GeoTIFF
│   │   ├── indices.py                  ← spectral index functions + INDEX_CALCULATORS registry
│   │   ├── normalization.py            ← per-channel mean/std normalization
│   │   ├── slope.py                    ← Horn slope calculation (windowed)
│   │   └── static_layers.py           ← warp / merge / fill substrate & bathy rasters
│   │
│   ├── model/                          ← model definition and weight loading
│   │   ├── __init__.py
│   │   ├── architecture.py             ← SegModel (PyTorch-Lightning U-Net + MaxViT-Tiny)
│   │   └── loader.py                   ← download weights from HuggingFace, return SegModel
│   │
│   ├── inference/                      ← tile-based inference engine
│   │   ├── __init__.py
│   │   └── engine.py                   ← DatasetInference: tiling, weighted stitching, save
│   │
│   ├── masking/                        ← post-inference spatial filters
│   │   └── __init__.py                 ← exclusion zones, depth mask, eelgrass mask
│   │
│   ├── postprocessing/                 ← output creation
│   │   ├── __init__.py
│   │   └── mosaic.py                   ← maximum-value BC Albers mosaic from multiple scenes
│   │
│   └── static/                         ← packaged geodata (read-only)
│       ├── __init__.py
│       ├── masks/
│       │   ├── valid_depth_zone.gpkg
│       │   ├── BCMCA_ECO_VascPlants_Eelgrass_Polygons_DATA.shp
│       │   └── (associated .dbf, .shx, .prj files)
│       └── bathy_substrate/
│           ├── __init__.py
│           ├── Bathymetry.tif          ← place downloaded static files here
│           ├── Slope.tif
│           ├── NCC_substrate_20m.tif
│           ├── SOG_substrate_20m.tif
│           ├── WCVI_substrate_20m.tif
│           ├── QCS_substrate_20m.tif
│           ├── HG_substrate_20m.tif
│           ├── BoPs_HG_10m.tif
│           ├── BoPs_NCC_10m.tif
│           ├── BoPs_QCSSOG_10m.tif
│           └── BoPs_WCVI_10m.tif
│
├── tests/                              ← pytest unit-test suite (no GPU / network required)
│   ├── __init__.py
│   ├── conftest.py                     ← shared fixtures
│   ├── test_indices.py                 ← spectral index functions
│   ├── test_normalization.py           ← mean/std normalization
│   ├── test_slope.py                   ← Horn slope calculation
│   ├── test_static_layers.py           ← fill_nodata and substrate helpers
│   ├── test_inference_engine.py        ← weight map, tile generation, index computation
│   └── test_model_loader.py            ← SegModel instantiation and load_model (mocked)
│
├── notebooks/
│   └── rasterizeNearshoreBottomPatches_BoPs.ipynb
│
├── pyproject.toml                      ← build config + dev extras (pytest, black, flake8)
└── README.md

---

📜 License

• Code: MIT License (see LICENSE file)
• Model: The trained model is licensed under CC-BY-4.0 — please cite the DOI when using it.