Running inference

We can use the trained segmentation models to extract site boundaries from satellite imagery.

The retrieve_shape_for_site.py script is a helper that:

1. Converts the given coordinates to a square polygon, add some extra padding.
2. Retrieve satellite imagery from esri or naip and assembles a GeoTIFF.
3. Runs the segmentation model, extracting all site features in the image.
4. Converts the mask into a GeoJSON (WGS84), ready to use in the platform.

As long as the tooling is set up correctly (Tools), it's straightforward to run the script.

1. Download the model weights and place them in models/.
2. Run the script, adjusting parameters as needed.

python scripts/retrieve_shape_for_site.py --model-weights models/model_final.pth --longitude "-103.73360747583735" --latitude "32.422276756574455" --provider esri

Parameters

Parameter	Type	Default	Description
--provider	string	"naip"	Image provider to use. Options: "esri" or "naip". NAIP provides aerial imagery for the US, while ESRI provides global satellite imagery.
--model-weights	string	Required	Path to the trained model weights file. Must be compatible with Detectron2 Mask R-CNN architecture.
--padding	float	60.0	Buffer distance in meters around the specified coordinates. Controls the size of the area to retrieve and analyze.
--cut-off	float	0.7	Confidence threshold for model predictions. Higher values result in fewer but more confident detections.
--latitude	float	Required	Latitude coordinate (in decimal degrees) of the target location.
--longitude	float	Required	Longitude coordinate (in decimal degrees) of the target location.
--output-folder	string	"./output"	Directory where output files will be saved. Will be created if it doesn't exist.
--zoom	integer	Provider-dependent	Zoom level for imagery retrieval. If not specified, defaults to 19 for ESRI and 18 for NAIP. Higher values provide more detailed imagery but cover smaller areas.
--device	string	"cpu"	Device to run the model on. Options: "cpu" or "cuda". Use "cuda" for GPU acceleration if available.

Outputs

File	Location	Description
{base_name}.geojson	Output folder	Final GeoJSON output containing detected features with confidence scores
{base_name}.tif	_cache subfolder	Cached GeoTIFF image with geospatial information
{base_name}.jpg	_cache subfolder	JPG version of the image used for model input
{base_name}.geojson	_cache subfolder	Temporary GeoJSON file used for image retrieval