Add MMFlood dataset

[lccol]

This PR adds the MMFlood dataset from the paper "MMFlood: A Multimodal Dataset for Flood Delineation From Satellite Imagery". This is a Sentinel-1 + DEM dataset for Image Segmentation.

Original tif files are of variable resolution. Max height in pixels is 2147, max width in pixels is 2313 (which are the ones reported in the docs). The dataset also includes hydrography information, but it is not available for all acquisitions (currently the implemented class does not read such tif files).

Example with False Color representation

[lccol]

@microsoft-github-policy-service agree

[nilsleh]

Thanks so much for the contribution, this is great! I have made a first pass of comments below. If you have questions about anything feel free to comment.

[adamjstewart]

Mostly looks good now, thanks for the hard work!

Only other comment I would make is that the recommended approach for these "curated" geospatial datasets (RasterDatasets containing both images and masks) is to create a dummy dataset for the images, a dummy dataset for the masks, and an IntersectionDataset that combines them. This usually lets you completely skip the __init__ and __getitem__ since it will inherit from RasterDataset. See L7Irish and L8Biome for examples of these. Up to you whether or not you want to do this since you're almost done, but it could make the code a bit cleaner.

[lccol]

Thank you @adamjstewart for your comments. You managed all of your comments, including the conversion of MMFlood to RasterDataset class, similarly to L7Irish and L8Biome. I have just two questions:

• I noticed that few entries have missing values, leading to NaNs. I was thinking of either putting all pixel values and mask values to 0 (option 1) or add a new entry missing (option 2) in the dict returned by the __getitem__. It will contain a Tensor of the same shape as the mask, with True in case of missing values in the image, False otherwise. I tried to check but I haven't found any other dataset which have NaN values in them. Probably this last option is kind of unusual compared to all the other datasets within the library... What do you think?
• from some tests, I found that some of the tiles are partially overlapped and thus different tif files are merged when doing an iteration over the entire dataset. This should be fine I guess, since both images and masks are merged in a consistent manner (same order for both Sentinel-1 and masks data in the reverse painter algorithm). However, the dataset uses the tags in the tif files to store the timestamp of each tile. Is there a way to create the RTree with the temporal information stored in the tags? From my understanding, RasterDataset parses dates directly from the filename...

[adamjstewart]

• I definitely prefer option 1. Our trainers support ignore_index which can be used to ignore these values during performance computation.
• You mean the timestamp isn't stored in the filename, it's stored in some kind of metadata? How do you access this metadata? It's possible to override __init__ and extract the appropriate metadata yourself, it's just really ugly.

[lccol]

• I definitely prefer option 1. Our trainers support ignore_index which can be used to ignore these values during performance computation.

Ok. I should have implemented it using the ignore_index.

• You mean the timestamp isn't stored in the filename, it's stored in some kind of metadata? How do you access this metadata? It's possible to override __init__ and extract the appropriate metadata yourself, it's just really ugly.

Yes, that is correct. To access this timestamp I should do as follows

with rasterio.open(path, “r”) as src:
        timestamp = src.tags()[“event_date”]

I haven’t checked properly, but I think that should not be a big issue the current way it is implemented (i.e. do not retrieve the timestamp for each date), since I believe most of the overlapping tiles refers to the same date…

[adamjstewart]

More comments to try to make the code simpler