[5pt] Misalignment between the input DEM and inundation predictions

[AliForghani-NOAA]

The DEM files are combined into a VRT to include a buffer for each HUC8. However, clipping the VRT to the extent of the buffered HUC8 results in a misaligned DEM. This misalignment propagates into the REM and inundation predictions.

The plot below shows the extent of the original DEM (black boxes) compared to the colormap of the clipped DEM. Additionally, using the GDAL argument -r bilinear causes the clipped DEM values to be inconsistent with the original DEM, further complicating debugging efforts.

The following code can be used to replicate the observations for HUC8 19020302:

gdalwarp -cutline data/inputs/pre_clip_huc8/20241002/19020302/wbd_buffered.gpkg -crop_to_cutline   -ot Float32   -r bilinear  -of "GTiff"  -overwrite 
 -co "BLOCKXSIZE=512"  -co "BLOCKYSIZE=512"  -co "TILED=YES"  -co "COMPRESS=LZW"   -co "BIGTIFF=YES"  -t_srs 'EPSG:3338'  -tr 10 10 data/inputs/dems/3dep_dems/10m_South_Alaska/20240912/FIM_3dep_dem_South_Alaska_10m.vrt  
 outputs/dem_meters_19020302.tif

[RobHanna-NOAA]

Could this be related to our system being in 10m rez? I wondered if this might popup.

Either in pre-clip or before, we need to:

1. Ensure all incoming data inputs are either 5070 or 3338 (AK) before any pre-clips.
2. Resolution has to be the same: ie) if we have any 3m DEM's coming in, they need to be adjusted back to 10m before clipping.

Ali: Could this be the case here? I suspected we had a mismatch in AK for the two conditions above.

[AliForghani-NOAA]

Rob, the misalignment only occurs when clipping the VRT (from a 10m DEM) into a 10m clipped buffered DEM. This issue is present for both CONUS and Alaska HUCs. Below is a plot for CONUS HUC8 02050206, where the gray-scale pixels represent the clipped DEM and the red pixel represents the original DEM. I don't see that pre-clipping is contributing to this issue.

[AliForghani-NOAA]

This issue also happens for DEM difference raster creation. Note that both clipped rasters are misaligned compared with the original raster.

This change is significant for the example shown here . In this case, using the bilinear method would show only the middle cell as non-inundated. However, switching to the near method makes all three bridge cells non-inundated. This occurs because the bilinear interpolation reduces the elevation differences by "smoothing" them, using neighboring non-bridge cells with zero difference values.

Because of this issue, and until we fix the misalignment, I am using -r near argument in clipping the difference rasters.