Skip to content

Processing Images

Run segmentation inference on your geospatial imagery using the hab segment command.

Prerequisites

  1. Activate Environment: Ensure (habitat-env) is active in your terminal.
  2. Input Data: Have your GeoTIFF orthomosaics or raster data ready.

1. Select a Model

Identify the appropriate model for your sensor capabilities and target species.

hab models
Model Name Revision Description Status
kelp-ps8b 20250818 Kelp segmentation model for 8-band PlanetScope imagery. Available
kelp-rgb 20240722 Kelp segmentation model for RGB drone imagery. Available
kelp-rgbi 20231214 Kelp segmentation model for 4-band RGB+NIR drone imagery. Available
mussel-gooseneck-rgb 20250725 Mussel and gooseneck barnacle segmentation model for RGB drone imagery. Available
mussel-rgb 20250711 Mussel segmentation model for RGB drone imagery. Available

2. Command Interface

The hab segment command requires the model name, input path, and output destination.

Syntax

hab segment --model <NAME> --input <PATH> --output <PATH> [OPTIONS]

For Beginners

You only need three things to get started:

  • --model or -m - Which model to use (e.g., kelp-rgb)
  • --input or -i - Path to your image file
  • --output or -o - Path where results will be saved

Advanced options below are for optimization and special cases. You can safely ignore them for your first run.

Parameters

Flag Short Description
--model -m (Required) Model identifier (e.g., kelp-rgb).
--input -i (Required) Path to source raster (.tif).
--output -o (Required) Path for output classification raster (.tif).
--crop-size -z Tile size for inference window (pixels). Default: 1024.
--batch-size Inference batch size. Increase for GPU acceleration. Default: 1.
--band -b Band mapping for non-standard sensor ordering (see below).
--blur Kernel size for median blur post-processing. Default: 5.
--morph Kernel size for morphological opening/closing. Default: 0.

Full Command Reference

For complete documentation of all available flags and options, see the CLI Reference.

3. Execution Examples

Standard RGB Inference

Process a standard RGB orthomosaic for kelp detection.

hab segment -m kelp-rgb -i "site_01.tif" -o "site_01_class.tif"

hab segment -m kelp-rgb -i "site_01.tif" -o "site_01_class.tif"

Process

The CLI will tile the input raster, perform inference, stitch the results, and apply geospatial metadata to the output file.

4. Advanced Configuration

Memory Management (OOM Errors)

For high-resolution rasters on machines with limited VRAM/RAM, reduce the tile size.

hab segment -m kelp-rgb -z 512 -i input.tif -o output.tif
(Note: crop-size must be an even integer).

Band Mapping

Models expect specific input band orders (e.g., kelp-rgbi expects Red, Green, Blue, NIR). If your sensor outputs a different order (e.g., Blue, Green, Red, NIR), use -b flags to map source indices (1-based) to the expected model input slots.

Example:

  • Source: Blue (1), Green (2), Red (3), NIR (4)
  • Model Expectation: Red, Green, Blue, NIR

Command: 

# Map Source Band 3 -> Slot 1 (Red)
# Map Source Band 2 -> Slot 2 (Green)
# Map Source Band 1 -> Slot 3 (Blue)
# Map Source Band 4 -> Slot 4 (NIR)
hab segment -m kelp-rgbi -b 3 -b 2 -b 1 -b 4 -i input.tif -o output.tif

5. Troubleshooting

Error Cause Solution
command not found Environment inactive Activate virtual env: source habitat-env/bin/activate or .\Scripts\activate.
CUDA out of memory VRAM limit exceeded Reduce --crop-size to 512.
Permission denied File lock Close the raster in GIS software (ArcGIS/QGIS) before processing.

Something else? Check out the FAQs

Next Steps

Post-Processing Methodology