Broadcast error with mosaic_reader past the bounds of a dataset

Hi 👋

It seems like attempting to build a mosaic with a bbox that extends beyond the bounds of a dataset creates an inconsistent data shape between datasets. I think this is happening because WarpedVRT reader doesn’t support boundless reads, so the shape of the data is clipped to the dataset’s bounds? I’ve worked around this by using a custom reader that uses rasterio.wrap.reproject after reading from the COG directly (instead of reprojecting on the fly).

I put together a short script demonstrating this issue with a couple of NAIP COGs from Planetary Computer:

import rio_tiler

cog_urls = [
    'https://naipeuwest.blob.core.windows.net/naip/v002/ca/2020/ca_060cm_2020/35116/m_3511663_se_11_060_20200523.tif',
    'https://naipeuwest.blob.core.windows.net/naip/v002/ca/2020/ca_060cm_2020/35116/m_3511663_ne_11_060_20200523.tif',
]

bbox = [-116.15799958899584, 35.059243863202646, -116.15220014855299, 35.06651998829758]

def _reader(asset, *args, **kwargs):
  with rio_tiler.io.COGReader(asset) as cog:
    return cog.part(*args, **kwargs)

image_data, _assets = rio_tiler.mosaic.mosaic_reader(
    cog_urls,
    _reader,
    bbox,
)

COG & BBox Geometry

{
  "type": "FeatureCollection",
  "features": [
    {
      "type": "Feature",
      "geometry": {
        "type": "Polygon",
        "coordinates": [
          [
            [
              -116.1522,
              35.059244
            ],
            [
              -116.158,
              35.059244
            ],
            [
              -116.158,
              35.06652
            ],
            [
              -116.1522,
              35.06652
            ],
            [
              -116.1522,
              35.059244
            ]
          ]
        ]
      },
      "properties": {
        "name": "read bounds",
        "fill": "green"
      }
    },
    {
      "type": "Feature",
      "geometry": {
        "type": "Polygon",
        "coordinates": [
          [
            [
              -116.122847,
              35.060229
            ],
            [
              -116.122135,
              35.126769
            ],
            [
              -116.189694,
              35.12724
            ],
            [
              -116.190352,
              35.060699
            ],
            [
              -116.122847,
              35.060229
            ]
          ]
        ]
      },
      "properties": {
        "cog url": "https://naipeuwest.blob.core.windows.net/naip/v002/ca/2020/ca_060cm_2020/35116/m_3511663_ne_11_060_20200523.tif"
      }
    },
    {
      "type": "Feature",
      "geometry": {
        "type": "Polygon",
        "coordinates": [
          [
            [
              -116.122792,
              34.997741
            ],
            [
              -116.12208,
              35.064281
            ],
            [
              -116.189654,
              35.064752
            ],
            [
              -116.19031,
              34.99821
            ],
            [
              -116.122792,
              34.997741
            ]
          ]
        ]
      },
      "properties": {
        "cog url": "https://naipeuwest.blob.core.windows.net/naip/v002/ca/2020/ca_060cm_2020/35116/m_3511663_se_11_060_20200523.tif"
      }
    }
  ]
}

</details

Traceback:

Traceback (most recent call last):
  File "rio-tiler-boundless.py", line 15, in <module>
    image_data, _assets = rio_tiler.mosaic.mosaic_reader(
  File ".local/lib/python3.9/site-packages/rio_tiler/mosaic/reader.py", line 94, in mosaic_reader
    pixel_selection.feed(img.as_masked())
  File ".local/lib/python3.9/site-packages/rio_tiler/mosaic/methods/defaults.py", line 20, in feed
    pidex = self.tile.mask & ~tile.mask
ValueError: operands could not be broadcast together with shapes (4,1232,982) (4,1231,981) 

Not totally sure if I’ve understood the cause here correctly, or of how to fix this in rio-tiler, but I’m happy to contribute a fix with guidance if you’re interested. Thanks!

[geospatial-jeff]

because looking at the geo info from the 2 COG their resolution is the same!

The resolution in native CRS is the same but are slightly different after being projected to match the CRS of the bounding box (4326) when calculating height/width etc. for the VRT. Both images are from the same quadrangle (ne/se) and there is latitudinal distortion projecting from NAD83 UTM to WGS84. If the two images you’ve selected were east/west oriented (ex. ne/nw or se/sw) you probably wouldn’t see this issue.

The resolutions of the two images (in degrees) after reprojection are:

5.909249718773602e-06
5.907606274181144e-06

Setting width and height should fix this; another workaround is to reproject the bbox to the native CRS of the image which will avoid the rounding error when reprojecting to WGS84:

import rio_tiler
from rasterio.crs import CRS
from rasterio.warp import transform_bounds

cog_urls = [
    'https://naipeuwest.blob.core.windows.net/naip/v002/ca/2020/ca_060cm_2020/35116/m_3511663_se_11_060_20200523.tif',
    'https://naipeuwest.blob.core.windows.net/naip/v002/ca/2020/ca_060cm_2020/35116/m_3511663_ne_11_060_20200523.tif',
]

bbox = [-116.15799958899584, 35.059243863202646, -116.15220014855299, 35.06651998829758]
bbox26911 = transform_bounds(
    CRS.from_epsg(4326), CRS.from_epsg(26911), *bbox
)

def _reader(asset, *args, **kwargs):
  with rio_tiler.io.COGReader(asset) as cog:
    return cog.part(*args, **kwargs)

image_data, _assets = rio_tiler.mosaic.mosaic_reader(
    cog_urls,
    _reader,
    bbox26911,
    bounds_crs=CRS.from_epsg(26911),
)

[lorenzori]

you are right, when I cropped the the tiles to share I messed with the resolution. Indeed if I project the originals to 3857 and assert they have the same resolution, the error from mosaic_reader disappears. Thank you!