Adding OSM data masks to an existing torchGeo geospatial data set

[dsterlingdg]

I would like to create a custom data loader that wraps an existing torchGeo geospatial data set to return a data mask containing Open Street Maps (OSM) labels in addition to the default image and label mask. I'm relatively new to working with geospatial data --- are there primitives in torchGeo that I can/should use to carry out the task of acquiring the OSM data for a given bounding box and then performing any coordinate system transformation/registration tasks required to get pixel level alignment of OSM labels and the raster data?

[calebrob6]

Hi @dsterlingdg, this doesn't exist in TorchGeo (although would be a cool feature!). To implement this you would need to know how to talk to an OSM API (e.g. Overpass). The osmnx library does this and I have an example notebook that uses osmnx to rasterize roads networks from that to pixel-alignment with other imagery -- https://gist.github.com/calebrob6/5039fd409921606aa5843f8fec038c03.

[dsterlingdg]

I expected to need to learn my way around a suitable OSM API, thanks for sharing your example @calebrob6! Let me play with this for a bit and I'll reach out if I have additional questions

[adamjstewart]

To add to what @calebrob6 said, you would want to create a new OSM dataset that subclasses GeoDataset and implement the __getitem__ method that creates a raster given a geospatial bounding box. All other stuff (wrapping existing datasets, adding to images and label masks, CRS transforms, pixel alignment, etc.) will be taken care of automatically by TorchGeo. Once OSM is written, the rest of the code would look like:

image = ...  # some image dataset, might be different res or CRS
mask = ...  # some mask dataset, might be different res or CRS
osm = OSM(...)

dataset = image & mask & osm
dataloader = DataLoader(dataset, ...)

[dsterlingdg]

Thanks @adamjstewart -- this was exactly what I had in mind. Let me see if I can cobble together the innards of an OSM dataset object.