Relation of geospatial machine learning (GeoML) data science libraries

[weiji14]

Inspired by @rbavery's blog post at https://developmentseed.org/blog/2022-09-27-ml-tooling, here's a table of GeoML Python libraries/frameworks focusing on the Data Ops side. This is a living document, so let me know if you see anything that needs updating!

	zen3geo	torchgeo	eo-learn	raster-vision
Year started	2022	2021	2018	2018
Organization	-	Microsoft	Sentinel-Hub/Sinergise -> Planet	Azavea -> Element84
Version	v0.6.1	v0.4.1	v1.4.1 (production)	v0.20.2
Design	Lightweight library	Heavyweight library	Library/Framework	Library/Framework
Paradigm	Composition	Inheritance	Inheritance	Chained-inheritance
Data model	xarray/geopandas	GeoDataset (numpy/shapely)	EOPatch (numpy/geopandas)	Scene (numpy/geopandas)
STAC	Yes	No^	No	Yes
Segmentation	Yes	Yes	Yes	Yes
Object Detection (bbox)	Yes	DIY	No?	Yes
Multi-CRS without reprojection	Yes	No	No?	No?
Multi-resolution	Yes via datatree + some DIY	No	Yes	No
Multi-dimensional (>2D+bands)	Yes, via xarray	No	No	No
Visualization	No	Yes	Yes	Yes
BYO custom function	Easy	Hard	Hard	Hard
ML Library coupling	Pytorch or None^	Pytorch	None	Pytorch, Tensorflow (DIY)

Key:

• ^: coming
• DIY - Do it Yourself

Tying 'features' to 'applications':

• Multi-resolution - enables super-resolution, handle multimodal sensors (optical/SAR/gravity/etc)
• Multi-dimensional - enables time-series analysis, forecasting with physics-informed neural networks (climate models, hazard predictions) and beyond

Note that these extra 'features' come for free by following the xarray Data Model.

References:

• PyPI GIS/Image Processing libraries
• PyPI GIS/Artificial Intelligence libraries