Compositors.md

Compositors

This document lists compositors available for raster layers in rslearn. Set compositing_method in a layer config to either:

• a built-in enum name (for example "FIRST_VALID"), or
• a custom compositor config (class_path + init_args).

See Dataset Configuration for where this field appears.

Built-In Compositors

Compositing Method	Notes
FIRST_VALID	Default. First non-nodata pixel in item order.
MEAN	Per-pixel mean across valid pixels.
MEDIAN	Per-pixel median across valid pixels.
SPATIAL_MOSAIC_TEMPORAL_STACK	Spatial first-valid per timestep, stacked along T.
TEMPORAL_MEAN	Temporal reduction over stacked timesteps.
TEMPORAL_MAX	Temporal reduction over stacked timesteps.
TEMPORAL_MIN	Temporal reduction over stacked timesteps.

Custom Cloud-Aware Compositors

These compositors use cloud or classification signals inside each materialized item group to rank or select items.

Class Path	Description
rslearn.dataset.hls_fmask.HlsFmaskFirstValid	Uses HLS Fmask QA bits to score cloudiness.
rslearn.dataset.omni_cloud_mask.OmniCloudMaskFirstValid	Uses OmniCloudMask model inference on R/G/NIR.
rslearn.dataset.sentinel2_scl.Sentinel2SCLBestClear	Selects the single Sentinel-2 item with highest SCL clear cover.
rslearn.dataset.sentinel2_scl.Sentinel2SCLFirstValid	Uses Sentinel-2 SCL classes to score cloudiness.