fix statistics for coverage #684
Conversation
rio_tiler/utils.py
Outdated
| ) -> float: | ||
| i = numpy.argsort(values) | ||
| c = numpy.cumsum(weights[i]) | ||
| return values[i[numpy.searchsorted(c, numpy.array(quantiles) * c[-1])]] |
There was a problem hiding this comment.
will be removed with numpy 2.0
vincentsarago
force-pushed
the
patch/better-statistics-for-coverage
branch
from
b1e8d73 to
699a902
Compare
| # Avoid non masked nan/inf values | ||
| numpy.ma.fix_invalid(data, copy=False) | ||
|
|
||
| for b in range(data.shape[0]): | ||
| keys, counts = numpy.unique(data[b].compressed(), return_counts=True) | ||
| data_comp = data[b].compressed() |
There was a problem hiding this comment.
data[b].compressed() was called multiple times
| # Population standard deviation of cell values, taking into account coverage fraction. | ||
| "std": _weighted_stdev(data_comp, masked_coverage.compressed()), | ||
| # Median value of cells, weighted by the percent of each cell that is covered. | ||
| "median": _weighted_quantiles(data_comp, masked_coverage.compressed()), |
There was a problem hiding this comment.
std and median are now weighted by the coverage array
| assert stats[0]["count"] == 1.75 | ||
| assert stats[0]["median"] == 3 # 2 in exactextract |
There was a problem hiding this comment.
I have no idea why median gives a different results. I've tested a new numpy 2.0 method and it gives 3 while exactextract give 2. I don't want to over engineer the median calculation
There was a problem hiding this comment.
For a raster of type T exactextract is returning type T for the quantile and median calculations. T here is int64, so the median of 2.5 is getting truncated to 2. Maybe quantile/median should be returning float64 instead.
There was a problem hiding this comment.
oO that makes sense 🙏 thanks for having a look
| assert stats[0]["max"] == 9 | ||
| # exactextract takes coverage into account, we don't | ||
| assert stats[0]["minority"] == 1 # 1 in exactextract | ||
| assert stats[0]["majority"] == 1 # 5 in exactextract |
There was a problem hiding this comment.
minority and majority do not take coverage into account. We might do this later if needed
|
Note: I've been working on making exactextract available on pypi so we can integrate into our CI isciences/exactextract#87 |
closes #680
overtake #681
Better take
coverageinto account. This PR tries to matchexactextractresults!cc @j08lue