Spatial Shift and Inaccuracy in Global Buildings Dataset Compared to US Building Footprints #97
Description
There is a noticeable spatial shift and reduced positional accuracy in the global buildings dataset (e.g., for Kenya) compared to the US-specific Microsoft building footprints. This results in misaligned geometries when overlaid on high-resolution basemaps or reference layers, suggesting systematic geographic offsets.
Observed Behavior:
Global buildings data (e.g., from https://minedbuildings.z5.web.core.windows.net/global-buildings/...) often appear slightly shifted (typically a few meters).
US building footprints (e.g., from Microsoft’s open dataset) are much more precisely aligned with base imagery.
Coordinate precision (even at 12 decimal places) is not sufficient to overcome the alignment discrepancy, indicating the issue lies in the data source or model inference rather than formatting.
Expected Behavior:
Buildings should align closely with actual rooftop outlines in satellite or aerial imagery, as is the case with the US dataset.
Possible Causes to Investigate:
Use of different satellite imagery sources or resolutions between global vs. US models.
Lack of high-resolution ground truth or calibration data outside the US.
Differences in post-processing pipelines, projection systems, or georeferencing methods.
Simplified or downsampled geometry in global outputs for bandwidth/storage reasons.
Questions:
What alignment or correction methods were applied to the US building footprint dataset that are missing or different in the global pipeline?
Are there plans to reprocess or improve positional accuracy in global tiles using better base imagery or alignment techniques?
Is the global model trained with lower resolution or general-purpose imagery compared to the high-quality US imagery?