Orfeo backscatter: preserve sparse partitioner

[jdries]

The V2 code path of orfeo backscatter results in an RDD with a default partitioner:

openeo-geopyspark-driver/openeogeotrellis/collections/s1backscatter_orfeo.py

Line 1031 in 85ac055

tile_layer = geopyspark.TiledRasterLayer.from_numpy_rdd(

As a result, point extraction jobs suffer from bad performance, which would not happen if sparse partitioners are available.

The V1 code path does have some support for sparse partitioners:

openeo-geopyspark-driver/openeogeotrellis/collections/s1backscatter_orfeo.py

Line 609 in 85ac055

result = p.applySparseSpacetimePartitioner(tile_layer.srdd.rdd(),

[VictorVerhaert]

This might also explain why the point extractions for EUGW cost more than anticipated

[jdries]

The fix is available on staging. For the weed job, costs went from ~900 to ~100.

It is the aggregate temporal step for sentinel-1 which is now more effective:
BEFORE:
aggregate_temporal results in 53280 keys, using partitioner index: org.openeo.geotrelliscommon.package$SpaceTimeByMonthPartitioner$@3a993035 with bounds KeyBounds(SpaceTimeKey(0,0,1704067200000),SpaceTimeKey(36,39,1734739200000))

AFTER:

aggregate_temporal results in 2628 keys, using partitioner index: SparseSpaceTimePartitioner 1620 true with bounds KeyBounds(SpaceTimeKey(0,0,1704067200000),SpaceTimeKey(36,39,1734739200000))

This is probably caused by the fact that aggregate temporal will insert empty tiles where it expects data.

[VictorVerhaert]

So if I understand correctly this influences PG's that use sar_backscatter followed by an aggregate_temporal? (so not only point extractions)