Radar Smoothing

Radars capture observations at different gate lengths, which are then processed into a final product that looks pixelated. How do we then take each “pixel” and create a smoothed radar product? Here’s an example of a source radar product and a smoothed version from GRLevelX.

radar smoothing from GRLevelX - https://www.grlevelx.com/

GR3 uses bilinear interpolation

superres l2 … bins have such a poor aspect ratio that even bilinear interpolation fails at reasonable ranges

What Gr3 Does

GR3 uses bilinear interpolation so the dbz at the center of a smoothed bin should be (almost) exactly the bin’s value. I don’t think there’s any doubt that the smoothed display more closely matches what’s out in the real world … assuming that the radar isn’t grossly under sampling the phenomena.

Superres L2 data is another matter! The bins have such a poor aspect ratio that even bilinear interpolation fails at reasonable ranges, ie. the bilinear kernel is exposed on the display. In addition, there appears to be more noise in the data due to fewer samples being used in a bin. I struggled with this for a while and looked at various techniques. I went as far as implementing the nexrad recombination algorithm in a pixel shader but the results came back as expected: a bloated/blobby/unnatural mess. In the end, I came up with what I believe is an acceptable solution in GR2Analyst. It performs a range-dependent weighted moving average based on the aspect ratio of the bins. All of this is done in the pixel shader while drawing.

About the gradients in Jeff’s post. It’s not the gradient that appears large between X=5 and X=6, it’s the delta (aka. discrete) difference. The gradient in the averaged bins (delta / distance) is actually lower than in the higher res bins, as expected.

Edit: IMHO, Superres data looks terrible zoomed out, whether it’s smoothed or not, because the data is noisy and badly undersampled on typical monitors. However, I may have stumbled upon a pretty good solution while experimenting with smoothing superres data.




Related Notes