support H3 spatial indexed data via the h3j extension?

[cboettig]

Would you be open to a PR adding support for H3-indexed data?

I think the simplest approach would use h3j, as linked from the maplibre plugin page.

Motivation: H3 is a powerful hierachical spatial index. Because it is supported through a duckdb extension, R users can leverage dplyr to efficiently calculate and manipulate h3-indices, e.g. in group-by/summarise aggregations and joins.

[cboettig]

I sketched this out in my fork by adding a add_h3j_sources() method that plugs into the existing layer interface like so

My approach adds a new function add_h3j_source mirrors the way the javascript extension provides addH3JSource() (unlike the PMTiles extension which took a different route but introducing a custom "protocol" for pmtiles:// layerd over the add_vector_source().

url = "https://inspide.github.io/h3j-h3t/examples/h3j/sample.h3j"
maplibre(center=c(-3.704, 40.417), zoom=15, pitch=30) |>
add_h3j_source("h3j_testsource",
              url = url
)  |>
  add_fill_extrusion_layer(
    id = "h3j_testlayer",
    source = "h3j_testsource",
    fill_extrusion_color = interpolate(
      column = "value",
      values = c(0, 21.864),
      stops = c("#430254", "#f83c70")
    ),
    fill_extrusion_height = list(
      "interpolate",
      list("linear"),
      list("zoom"),
      14,
      0,
      15.05,
      list("*", 10, list("get", "value"))
    ),
    fill_extrusion_opacity = 0.7
  )

[walkerke]

This looks awesome! Yes I'd be interested in a PR.

Motivation: H3 is a powerful hierachical spatial index. Because it is supported through a duckdb extension, R users can leverage dplyr to efficiently calculate and manipulate h3-indices, e.g. in group-by/summarise aggregations and joins.

I'm wondering, given your comment: would this be a method we could use for efficient plotting of high-density point data, e.g. bin the data and visualize without having to send a hefty GeoJSON over to be rendered?

[walkerke]

@cboettig one thing I notice is that the repo does not appear to be maintained (although it's good to know that it still works); perhaps we fork it and modify as needed - I saw your open issue. I also note that they are using a very old version of maplibre on their side, so we'd want to make sure it works with anything new.

[cboettig]

@walkerke Thanks Kyle! These are good questions.

Yeah I hear you on the maintenance side, that's a good question. It have it working fine in my fork, so at least so far seems good.

would this be a method we could use for efficient plotting of high-density point data, e.g. bin the data and visualize without having to send a hefty GeoJSON over to be rendered?

yes, this is precisely my use case. e.g. GBIF is a biodiversity occurrence dataset of some 3 billion rows, about half a terrabyte as parquet point data. Here is a demo app that bins the data by H3 hexes at various resolutions. You can aggregate data for the country or the world using low zooms (h4 or say) quite quickly, or zoom into a state or national park and aggregate at much finer resolutions: https://streamlit.carlboettiger.info (this one python/streamlit based, but we can do the same in R with your package now, and probably make it look much nicer :-) )

I should note that deck.gl also has support for h3, but that would not integrate into the existing maplibre syntax for add_fill_layer, add_extrusion_layer and all the power that comes with those, so a user would have to use a different function to tell deck.gl how to handle data color, fill, etc.

I don't know javascript but yeah I think we could fork it if necessary.

[cboettig]

@walkerke just wanted to check in if there's specific changes/additions you'd like to see here?

[walkerke]

no, I think it looks good! I'm going to merge. One thing we'll want to think about is how (or if) to support conversion of sf objects into the spec required by H3J. The biggest value-add would be for a user who has say 1 million points to map, and wants to use hexagonal binning to visualize that. We'd probably want to do this either through a dedicated conversion function, or with some documentation on how users can do the conversion themselves.