geodeep 0.9.5

Free and open source library for AI object detection in geospatial rasters

GeoDeep

A fast, easy to use, lightweight Python library for AI object detection in geospatial rasters (GeoTIFFs), with pre-built models included.

Install

pip install -U geodeep

Usage

From the command line

geodeep [geotiff] [model ID or path to ONNX model]

Example:

geodeep orthophoto.tif cars

Here GeoDeep will find cars in the orthophoto and write the result as a GeoJSON file containing the bounding boxes, confidence scores and class labels.

A list of up-to-date model IDs can be retrieved via:

geodeep --list-models

See also geodeep --help.

From Python

from geodeep import detect
bboxes, scores, classes = detect('orthophoto.tif', 'cars')
print(bboxes) # <-- [[x_min, y_min, x_max, y_max], [...]]
print(scores) # <-- [score, ...]
print(classes) # <-- [(id: int, label: str), ...]

geojson = detect('orthophoto.tif', 'cars', output_type="geojson")

Models by default will be cached in ~/.cache/geodeep. You can change that with:

from geodeep import models
models.cache_dir = "your/cache/path"

Models

Model	Description	Resolution (cm/px)	Experimental
cars	YOLOv7-m model for cars detection on aerial images. Based on ITCVD.	10
trees	Retinanet tree crown detection model from DeepForest	10	:heavy_check_mark:
trees_yolov7	YOLOv9 model for treetops detection on aerial images. Model is trained on a mix of publicly available datasets.	10	:heavy_check_mark:
birds	Retinanet bird detection model from DeepForest	2	:heavy_check_mark:
planes	YOLOv7 tiny model for object detection on satellite images. Based on the Airbus Aircraft Detection dataset.	70	:heavy_check_mark:

All ONNX models are published on https://huggingface.co/datasets/UAV4GEO/GeoDeep-Models

Creating New Models

Instructions coming soon. The basic idea is to create an ONNX model (see the retinanet conversion script) and possibly make some modifications to GeoDeep to handle different conventions in model architectures via conditional checking.

Why GeoDeep?

Compared to other software packages (e.g. Deepness), GeoDeep relies only on two dependencies, rasterio and onnxruntime. This makes it simple and lightweight.

Does this need a GPU?

It does not! Models are tuned to run fast on the CPU.

Contributing

We welcome contributions! Pull requests are welcome.

Roadmap Ideas

• Train more detection models
• Add support for semantic segmentation models
• Faster inference optimizations

Support the Project

There are many ways to contribute to the project:

• ⭐️ us on GitHub.
• Help us test the application.
• Become a contributor!

Credits

GeoDeep was inspired and uses some code from Deepness and DeepForest.

License

The code in this repository is licensed under the AGPLv3.

Made with ❤️ by UAV4GEO