Issue with GridGeoSampler in Windows 11

[FatemehEsfahani]

Description

I want to use GridGeosampler in Windows 11: sampler = GridGeoSampler(resulting_dataset, size=128, stride=128). However, nothing is produced. The length of sampler is zero.

Steps to reproduce

This is the bounding box of my resulting_dataset:
BoundingBox(minx=-120.03357866990702, maxx=-120.01546172995442, miny=49.703449529631186, maxy=49.71845099117338, mint=0.0, maxt=9.223372036854776e+18)
sampler = GridGeoSampler(resulting_dataset, size=128, stride=128)

Version

NA

[adamjstewart]

Difficult to reproduce without data or code. What's the image resolution?

[FatemehEsfahani]

Resolution is 30 meters

Difficult to reproduce without data or code. What's the image resolution?

[adamjstewart]

What's the resolution in the same units as the CRS? It appears to be in degrees. I'm trying to figure out how many 128x128 px patches should fit in your image.

[FatemehEsfahani]

What's the resolution in the same units as the CRS? It appears to be in degrees. I'm trying to figure out how many 128x128 px patches should fit in your image.

It is 0.00026949458523585647. CRS is EPSG:4326. I think the area is too small to samle from. So, in this case which sampler should we use to do padding if the area is small?

[adamjstewart]

So if I did my math correctly, you have approximately a single 56 x 67 px image? In that case, you don't really need a sampler at all, do you? You could just use:

sample = dataset[dataset.bounds]

to load the only image. You could also manually calculate a larger bounding box if you want to automatically pad to 128 x 128 px. Note that thousands of images are required to train a model, but you can still do inference on a single image if you already have a trained model.

[FatemehEsfahani]

So if I did my math correctly, you have approximately a single 56 x 67 px image? In that case, you don't really need a sampler at all, do you? You could just use:

sample = dataset[dataset.bounds]
to load the only image. You could also manually calculate a larger bounding box if you want to automatically pad to 128 x 128 px. Note that thousands of images are required to train a model, but you can still do inference on a single image if you already have a trained model.

I see. Thank you for your help.