Sampling

[m-mohr]

For several use cases such as ML training, one or more sampling processes could be useful.

This came up in the discussions around #295

Related work:

• https://eo-learn.readthedocs.io/en/latest/eolearn.geometry.sampling.html
• https://developers.google.com/earth-engine/apidocs/ee-image-sample?hl=en

Alternative:

• polygon_to_point?
• raster <-> vector processes #155

[clausmichele]

The sampling process could also allow to "extract" all the valid points setting a parameter. This would cover the case where we select some polygons (filter_spatial) and we want to keep all the pixels inside of them for training.

Some questions that come up in my mind:

1. Is the output a vector-cube?
2. Does it keep x,y information about the points or do we discard them?

[clausmichele]

The sampling will be performed using a new process called polygon_to_points:

• polygon_to_points takes as input a geoJSON (or other supported formats, TBD) and returns the same structure with point coordinates
• possible parameters are:

1. Sampling strategy: random, fibonacci, ...
2. Max number of points to extract
3. Other?

[m-mohr]

• We still need to verify which name to choose, they should be somewhat aligned between vector processes.
• Input should probably be GeoJSON or vector cube, as usual.
• Would it be useful to also pass lines and get points for them or should this be purely polygon focused?
• Sampling strategy could be a callback (so random and Fibonacci would be separate processes) or a given set of methods.

[m-mohr]

• Do we need more than random sampling? I couldn't really find (popular) implementations of Fibonacci.
• Number of samples: Allow percentage and an absolute number of pixels?

And lastly @clausmichele , I'm a bit confused about: "takes as input [...] and returns the same structure with point coordinates". This sounds like you'd extract points along the polygon borders, but I assume you meant that it samples from the whole inner polygon, right?

[clausmichele]

Sorry for being not so clear, with same structure I was referring to the geoJSON or vector-cube structure/data format, not the content itself.

Allow percentage and an absolute number of pixels?

I'm not sure actually. The percentage could be tricky since I would not know how to define a maximum number of points (100%) which could be extract from a vector layer, in theory they could be infinite!

[m-mohr]

Thanks for the clarification, @clausmichele.

I thought about basing percentages around the pixel centers. You should have a known number of pixels and could create a list of points for the pixel centers, right?

[clausmichele]

From the last meeting I understood that the process should take as input geometries (geoJSON) and output also geometries to be used in aggregate_spatial, that's why I firstly called it "polygon_to_points". The output would be a list of points which can be used in aggregate_spatial to create a vector-cube for training the ML model.

@mattia6690 @jdries can you comment on this? Do I remember correctly?

[m-mohr]

Ah, okay! Then I misunderstood or remembered incorrectly. I can change that. That makes the process simpler by only creating points from a shape without directly combining them with the values from the raster cube. That's also fine for me, just didn't fully understood the idea then. Stay tuned for an updated version... :-)

[ValentinaHutter]

From the last meeting I understood that the process should take as input geometries (geoJSON) and output also geometries to be used in aggregate_spatial, that's why I firstly called it "polygon_to_points". The output would be a list of points which can be used in aggregate_spatial to create a vector-cube for training the ML model.

@mattia6690 @jdries can you comment on this? Do I remember correctly?

I am working on an implementation of aggregate_spatial at EODC. Is there an example of aggregate_spatial being used with vector_to_random_points or vector_to_regular_points somewhere? Like an input of aggregate_spatial and the corresponding output? :)

[clausmichele]

This is my implementation of it https://github.com/SARScripts/openeo_odc_driver/blob/d1383aa872bdef5a8bde37f5d48b1f7a56cdd57e/openeo_odc_driver.py#L566

There's a lot of room for improvements, the properties are not kept for example. I will provide you also an example using vector_to_*_points soon.

[ValentinaHutter]

Thanks a lot, that's great! :)