[RFC] - Implement Aggregations on Geo Shape Field

[navneet1v]

The purpose of this RFC (request for comments) is to gather community feedback on a proposal to allow OpenSearch users to facilitate aggregations over the GeoShape data type.

Background

The geo_shape data type facilitates the indexing and searching with arbitrary geo shapes such as rectangles and polygons. It gets used when we want to index the shapes(like Line, MultiLine, Polygon etc) other than just points in a geographical context.

Aggregations

As of OpenSearch 1.0(derived from ElasticSearch 7.10.x), aggregations are supported on Geo Point data type not on the Geo Shape. We will be adding 4 different Aggregations on the Geo Shape data type of Open Search.

1. Geo Bounds Aggregation
2. Geo Centroid Aggregation
3. Geo Hash Aggregation
4. Geo Tile Aggregation

Geo Bounds Aggregation

A metric aggregation that computes the bounding box containing all geo values for a field.

Geo Centroid Aggregation

A metric aggregation that computes the centroid from all coordinate values for geo fields. The centroid calculation implements the ISO/IEC 13249-3:2016 SQL/MM specification, section 8.1.6.

Geo Hash Aggregation

Geo Hash is multi-bucket aggregation that will be grouping Geo Shapes into buckets which represent cells in a grid. Each cell is labeled using a geohash which is of user-definable precision. The internal working and output of the aggregation will be very similar to Geo Points except that a single shape can be counted for multiple tiles if any part of the shape intersects with that tile.

Geo Tile Aggregation

Geo Tile is multi-bucket aggregation that will be grouping Geo Shapes into buckets which represents map tiles as used by many online map sites. Each cell/tile will be labeled using the format : {zoom}/{x}/{y}, where zoom will be equal to the user-specified precision provided as input. The internal working and output of the aggregation will be very similar to Geo Points except that a single shape can be counted for multiple cells if any part of the shape intersects with that cell.

Plan

We will be launching the metrics aggregations(Geo Bounds and Geo Centroid) first and then moving towards the Geo Tile and Hash Aggregation.

Next Steps

As a next steps we will be adding 1 RFC for each of the above aggregation with more details around each aggregation.

Customer Request

1. https://forum.opensearch.org/t/support-for-geo-distance-queries-sorting-on-geo-shape-fields/7777/6
2. [FEATURE] Add support for geo_shape fields #55

References

1. GeoBound on Geo Points: https://opensearch.org/docs/latest/opensearch/metric-agg/#geo_bound
2. Zoom Level Documentation: https://wiki.openstreetmap.org/wiki/Zoom_levels
3. Centroid: https://en.wikipedia.org/wiki/Centroid
4. Geo Hash: https://en.wikipedia.org/wiki/Geohash
5. Map Tile: https://en.wikipedia.org/wiki/Tiled_web_map
6. ISO/IEC: https://www.iso.org/obp/ui/#iso:std:iso-iec:13249:-3:ed-5:v1:en

[AdrienF]

Hello, I just want to support this issue by explaining our use case, maybe it might help you prioritize it :

We want to build a "heatmap" of object detected on images. Objects are being modelized by axis aligned bounding boxes (stored as geo_shape of type Polygon or envelope) in our index, and scaled from normal coordinates to fit the Geo bounds [-180,180]x[-90,90].
Then, the heatmap is just a simple geotile_grid aggregation.

The problem comes from the fact that it is not yet possible in opensearch to aggregate on geo_shape field.

One workaround we currently have is to aggregate on detection centers, mapped as geo_point. It is not fully satisfactory as it does not take the detection area into account.

The ideal implementation for this feature would be an aggregation where geo_shape would be counted in each intersecting tile proportionally to its relative intersection area.

Additionnaly, the geo referential is useless to this application, do doing all this on arbitrary coordinates would be perfect.

Best regards.

Adrien

[navneet1v]

Hi Adrien,
Thanks for providing the detailed use case. These aggregations are prioritized in the roadmap and we are currently working on GeoBounds Aggregation(#93) to start with. We will keep this github issue updated for the release version we are targeting.

[navneet1v]

Uber level implementation Issue: #104