[Feature]: Identify land areas for hook & line survey when creating the grid

[chantelwetzel-noaa]

Is your feature request related to a problem? Please describe.
The NWFSC Hook & Line survey occurs in the Point Conception area sampling around multiple islands. These land masses need to be accounted for when developing the grid.

Describe the solution you'd like
Similar to the process currently in place for creating the grid for the NWFSC WCGBT survey (via the mask(bathy_raster, CCA) code), the land masses in the Hook & Line survey area need to be identified and removed from the grid.

Describe alternatives you've considered
None

Additional context
None

[kellijohnson-NOAA]

I believe that @jkbest2 has done this for islands in Alaska. I think that we would want to restrict the mesh but perhaps it might be that we need to modify the prediction grid as well. Any advice @seananderson, @ericward-noaa, @James-Thorson-NOAA, or @jkbest2?

[ericward-noaa]

For estimation, all you need to do is get a shapefile of the islands. Then it becomes easy to add a barrier mesh -- no further tweaks needed. Implemented in both VAST and sdmTMB.

If you're predicting to new locations -- then agree, you need to modify the prediction grid to not include islands.

[kellijohnson-NOAA]

Thanks @ericward-noaa I see an example in the {sdmTMB} documentation about adding a barrier mesh. Is this the best place to find information or should we be looking at stuff in the scratch pad for newer/different methods?

[James-Thorson-NOAA]

I could also put you in touch with Cecilia O'Leary and Arnaud Gruss, who've tested it in Aleutian Islands and New Zealand, respectively. They've found (using case studies and simulation experiments) that it doesn't make much difference in those systems. Arnaud is close to submitting a paper testing this (as a part of a larger study). In VAST it's just a matter of switching to Method = "Barrier", although it automatically turns off geometric anisotropy ... no one to my knowledge has derived how to do both jointly.

…

On Tue, Nov 15, 2022 at 8:23 AM Kelli Johnson ***@***.***> wrote: Thanks @ericward-noaa <https://github.com/ericward-noaa> I see an example in the {sdmTMB} documentation about adding a barrier mesh <https://pbs-assess.github.io/sdmTMB/reference/add_barrier_mesh.html>. Is this the best place to find information or should we be looking at stuff in the scratch pad for newer/different methods? — Reply to this email directly, view it on GitHub <#1 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AL62VMWQAQJJG6AQQXNC7XDWIO2IZANCNFSM5T4M6FXQ> . You are receiving this because you were mentioned.Message ID: ***@***.***>

[ericward-noaa]

No -- that's the best approach in sdmTMB right now. FWIW, the barrier mesh construction is an INLA thing -- so once you make it, it could be passed to whatever you want to use for estimation.

[kellijohnson-NOAA]

Thanks @James-Thorson-NOAA and @ericward-noaa. For the hook and line, turning off anisotropy might not be a big deal because it is constrained to a smaller footprint but for the WCGBTS survey, which is coastwide, I would think that we would be better off to just ignore islands 🤷 .

[James-Thorson-NOAA]

I'm willing to bet someone could derive the anisotropy + barrier feature ... to my knowledge the anisotropy feature isn't in R-INLA itself, and I just worked it out the matrix-construction equations from Lindgren-2011 (which we then merged up into TMB namespace R_inla) ... presumably a few days of reading and testing could yield the same thing with barrier feature too if someone is motivated

[ericward-noaa]

Also might be worth asking Ray Webster / IPHC if they've dealt with this before -- IPHC has been using barrier meshes for a while w/ custom code

[seananderson]

From my experience, the barrier rarely does much for index calculation—it's making subtle changes to the local distribution of fish and the land wasn't part of the prediction grid anyways. I'd guess that anisotropy is usually more important there. I think the place where the barrier is more useful is when the spatial distribution itself (i.e., nice maps) is the main interest. Even then, it seems like you need some fairly contrived examples to get strong effects. The main advantage I find is that it doesn't horrify ecologists when you tell them you're smoothing fish across land.