Improvements to RandomBatchGeoSampler

[tritolol]

I propose two improvements to RandomBatchGeoSampler:

1. The size parameter should be defined in pixels, not in CRS units. The output of a DataLoader using this sampler will be fed directly into a CNN or similar in most cases, which expect a certain defined input size in pixels. To accomplish this, the user has to perform the conversion of CRS to pixels manually like this

tile_size_pix = 256
sampler_size = tile_size_pix * input_ds.res
sampler = RandomBatchGeoSampler(input_ds, size=sampler_size, ...)

which is inconvenient.

2. The length parameter should have a meaningful default value. Currently, it is described as "length (int) – number of samples per epoch". Since an epoch is defined as processing an entire dataset exactly once, a meaningful default value could be calculated by using the dataset area and the sampler tile area

sampler_len = int(get_bbox_area(input_ds.bounds)/(sampler_size*sampler_size))
sampler = RandomBatchGeoSampler(input_ds, size=sampler_size, batch_size=8, length=sampler_len)

with

# This function would be nice to have in the BoundingBox class
def get_bbox_area(box):
    return (box.maxx - box.minx) * (box.maxy - box.miny)

Let me know if this makes sense.

[calebrob6]

Thanks @tritolol! Both changes make sense to me.

@RitwikGupta has something similar to your first change here https://github.com/microsoft/torchgeo/pull/294.

@adamjstewart is adding a volume attribute to BoundingBox, adding area in the same PR would make sense to me.

[TCherici]

With regards to point 1. by @tritolol, the actual output size of the sample mask is not guaranteed to be the actual number of pixels set by tile_size_pix (see #674). I am however assuming that this is a bug, and that it will be fixed soon