Feature/add preprocessing smoothing

UCL / pyCascadia

Implementation of GEBCO cookbook remove-restore and other cleaning of topography/bathymetry. Uses `pyGMT`.

Mozilla Public License 2.0

9 stars 0 forks source link

Feature/add preprocessing smoothing #45

Open alessandrofelder opened 3 years ago

alessandrofelder commented 3 years ago

This PR is an attempt at incorporating preprocessing ("workflow steps B+C of the GEBCO cookbook") into remove-restore.py. See #32 and #43 for details.

This PR mainly adds

a preprocessing function doing blockmedian, surface, grdfilter and grdsample on combined base and update grid data.
and incorporates this functionality into the main() remove-restore workflow

We need to think more about

how much this really helps us
if it does help, is there a better way to make the preprocessing part of the workflow?
- do we want to control this via input arguments?
- do we want a separate entry point for preprocessing?
- how should it work in the case when the user provides several input grids?

alessandrofelder commented 3 years ago

close-up-comparison This is a representative close-up comparison of using vs. not using the preprocessing steps B+C (in their current implementation). I ran:

remove-restore --base gebco_2020_n54.0_s43.0_w-130.0_e-117.0.nc --output=vancouver-island.nc --plot --spacing 0.00028 NOAA_South_VancouverIsland_30m.tif --region_of_interest -123.4 -123 47.8 48.0 --diff_threshold 0.1

It doesn't seem to make any real difference on the smoothness of the boundary? 🤔 For orientation, this is here:

D3varaj-G commented 3 years ago

Just to confirm: Input base grid is GEBCO ~300m Input update grid is for Vacouver ~30m Output spacing ~ 30m

Now: (Step B +C p. 148 in GEBCO Cookbook)

All input data combined and blockmedian filtered at 30m*4 = 120m grid
Blockmedianed data gridded at 120m using surface
Gridded data smoothed at 120m*3 = 3600m using grdflter
Filtered data resampled at 30m using grdsample (i.e. Step C of Cookbook) Output at this stage is value at nodes of 30m grid.

Next: (remove-restore)

All input data combined and blockmedian filtered at 30m grid (Output here is cell-centred!?)
Find difference between blockmedianed data and resampled data (i.e. output of Step B+C)
If difference above a threshold, replace node data in 30m grid with blockmedian data.

alessandrofelder commented 3 years ago

@Devaraj-G essentially, yes, that is what is happening.

The main difference between your description and what is happening in the code is that only the update grid gets put back into the output of Step B+C.

(small corrections, not important: 3*120=360, not 3600, and Step C is the grdfilter not the grdsample step)