Fix bucket assignment

[sfinkens]

The bucket assignment in the bucket resampler seems to be off by half a pixel. This PR provides a fix.

Example:

from pyresample import create_area_def
from pyresample.bucket import BucketResampler
import dask.array as da
import numpy as np

area = create_area_def(
    area_id='test',
    projection={'proj': 'latlong'},
    center=(0., 0.),
    width=4,
    height=4,
    resolution=10.0
)

lons = da.from_array(np.array([-15, -5, 5, 15]), chunks=2)
lats = da.from_array(np.array([-15, -5, 5, 15]), chunks=2)
data = da.from_array(np.array([1, 2, 3, 4]), chunks=2)
resampler = BucketResampler(source_lats=lats,
                            source_lons=lons,
                            target_area=area)
count = resampler.get_count()

Output with current master:

[[0 0 0 1]
 [0 2 0 0]
 [0 0 0 0]
 [1 0 0 0]]

Output with this PR:

[[0 0 0 1]
 [0 0 1 0]
 [0 1 0 0]
 [1 0 0 0]]

Changes:

• Compute bucket indices with respect to the edge of the corner
  pixel, not the center.

• Don't round projection coordinates. This is already done by
  casting to integer.

• That also means that we don't need the projection coordinates
  anymore. We can use the area extent instead.

• Tests added

• Tests passed

• Passes git diff origin/master **/*py | flake8 --diff

[mraspaud]

Where are we on this ?

[sfinkens]

@mraspaud This was just a quick hack to show @pnuu my findings. But neither @pnuu nor me really had the time to look into this in more detail. It's on my list and probably next week I will continue working on this.

[sfinkens]

Roughly one week later I finally managed to continue with this 😄

[coveralls]

Coverage increased (+0.03%) to 91.95% when pulling 577a2d1 on sfinkens:bucket-lowres into d8b1b88 on pytroll:master.

[codecov]

Codecov Report

Merging #249 into master will increase coverage by 0.02%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##           master     #249      +/-   ##
==========================================
+ Coverage   91.94%   91.97%   +0.02%     
==========================================
  Files          42       42              
  Lines        8098     8099       +1     
==========================================
+ Hits         7446     7449       +3     
+ Misses        652      650       -2     

Impacted Files	Coverage Δ
pyresample/bucket/__init__.py	98.91% <100.00%> (-0.06%)	⬇️
pyresample/test/test_bucket.py	100.00% <100.00%> (ø)
pyresample/geometry.py	82.50% <0.00%> (+0.09%)	⬆️
pyresample/area_config.py	87.74% <0.00%> (+0.27%)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update d8b1b88...577a2d1. Read the comment docs.

[sfinkens]

I found another little bug: Some observations outside the target area are assigned to the buckets in the first row/col. This happens if the observations are less than one bucket beyond the left/right boundary of the target area. In that case

-1 < (proj_x - adef.area_extent[0]) / x_res < 0
or
-1 < (adef.area_extent[3] - proj_y) / y_res < 0

so that .astype(int) yields 0 instead of -1. Flooring before casting to integer solves the problem.

[sfinkens]

Here is an example where I aggregated some GEO-LEO collocations to a regular grid and found too many observations in the first row and column:

[mraspaud]

Nice investigation work! tell us when you want us to review this :)

[sfinkens]

I think it's ready for review now :)

[sfinkens]

@ameraner Since you also started working with the bucket resampler, I'd be pleased if you could review this, too (if you have time)

[ameraner]

Looks good to me, as discussed in Slack :)

[sfinkens]

Thanks for taking the time @ameraner !

[pnuu]

LGTM