A little bit of feedback base on Swindon. See attached 1km2 screenshot from QGIS, including predicted road widths from LINDER:
This under predicts hard surfaces in two ways I can see:
Road widths are about right, but misses associated footpaths – could increase all widths by 2m or do separate calculation for paths assuming footpaths = roads+1m border.
Any road that is ‘clipped’ is excluded from the calculation. This makes sense for your larger examples like Columbo, but I wonder if it’s possible to clip shape based on border of box for these smaller examples?
For vegetation, here is a representative comparison b/w predicted and satellite (pretty close to mode prediction):
You can probably see this picks up large vegetation quite easily, but underpredicts veg in backyards/ streets (compare to QGIS image). I know the resolution in Sentinel won’t pick up all urban vegetation, but it looks like the algorithm is currently missing some veg pixels in sentinel. Is it possible to adjust the sensitivity to vegetation, or is this a hidden layer in the machine learning?
Based on @matlipson feedback.
A little bit of feedback base on Swindon. See attached 1km2 screenshot from QGIS, including predicted road widths from LINDER:
This under predicts hard surfaces in two ways I can see:
For vegetation, here is a representative comparison b/w predicted and satellite (pretty close to mode prediction):
You can probably see this picks up large vegetation quite easily, but underpredicts veg in backyards/ streets (compare to QGIS image). I know the resolution in Sentinel won’t pick up all urban vegetation, but it looks like the algorithm is currently missing some veg pixels in sentinel. Is it possible to adjust the sensitivity to vegetation, or is this a hidden layer in the machine learning?