Implement GeoPandasVectorSource

[JamesZDonline]

Description

While there are many potential vector sources out there, the only concrete implementation of the VectorSource class in RasterVision is the GeoJSONVectorSource. I'm proposing a GeoPandasVectorSource that takes a GeoPandas.DataFrame (or list of them) to construct a VectorSource.

Potential Advantages to a GeoPandasVectorSource

• Enhanced Flexibility: GeoPandas can read data from many different sources, (geopackages/shapefiles/PostGIS databases/etc.) Therefore, rather than having to convert data to geojson files before using them with RasterVision, we could use data from nearly any source directly.
• Simplified Workflows: GeoPandas has extensive tools for processing GIS data. So one could easily do a variety of transformations to the data prior to passing them to the machine learning algorithm without having to save copies of the transformed data.
• Reduced Storage Overhead: Minimizes the need to store multiple versions of data in GeoJSON format.

Would there be interest in this? Can anyone see any barriers or potential problems with this approach? From my understanding of the code, it seems like it would not be too challenging, but I'm happy to admit I've been wrong before and may have missed some crucial details. :)

We have a team member who is willing to take this on if there is interest.

Example use case:

Right now we have our vector data stored in a PostGIS database. When we want to run an experiment with a dataset, we connect to the database, pull it down as a GeoPandas.DataFrame, do a series of preprocessing steps (filtering, organizing, performing transformations, adding class_ids etc) before finally exporting them to geojson files (sorted in a directory structure by AOI and Label with a csv to match them up). If we want to run another experiment with slightly different data or different preprocessing, we run the whole process again and end up with another directory filled with several hundred geojsons. Tracking versioning has started to get complicated and confusing. If we could use the DataFrames directly, then we wouldn't need to store and version the whole data, we could just version the configuration of the preprocessing script and pull the data from our database again as needed.

[AdeelH]

I think this is a great idea and something I've thought of doing myself. Any contributions would be very welcome!

One wrinkle you might run into is that currently a lot of the GeoJSON-handling is actually part of the base VectorSource class rather than the GeoJSONVectorSource class, which is not great, so you might need to move that GeoJSON-specific stuff to GeoJSONVectorSource.

Please feel free to reach out if you have any other questions as you work on this!

[JamesZDonline]

Great! This is exciting. We've never contributed before, so we appreciate your support and patience. We'll get to work!

@simeonbetapudi We're good to go!

[AdeelH]

Wonderful! We will need you to sign a CLA though before we can accept any changes. You can find it here: https://docs.rastervision.io/en/stable/CONTRIBUTING.html.