Skip to content

Components

Each component in the pipeline is responsible for a single stage of processing. All components share the same interface: they receive a DataState, do their work, and return a ComponentResult.

All components can also be run individually — see Standalone Usage.

Tilerizer

Splits a large orthomosaic into smaller, overlapping tiles suitable for model inference.

Tile types:

Type Description
tile Unlabeled regular-grid tiles (input to detector or segmenter)
tile_labeled Labeled tiles with COCO annotations (input to prompted segmenter)
polygon Per-polygon tiles (input to classifier)

You can select the type of tilerizer you want using the tile_type parameter in your tilerizer config.

Requires: imagery_path (+ infer_gdf for tile_labeled and polygon)

Produces: tiles_path, optionally infer_coco_path


Detector

Runs object detection on image tiles, producing bounding box predictions with confidence scores.

Requires: tiles_path

Produces: infer_gdf (geometry, object_id, tile_path, score, class), infer_coco_path


Segmenter

Refines bounding box detections into instance segmentation masks. Can operate in prompted mode (using detector boxes) or unprompted mode.

Requires: tiles_path, optionally infer_coco_path (prompted mode)

Produces: infer_gdf with updated mask geometries, infer_coco_path


Aggregator

Merges overlapping detections from tiled inference using non-maximum suppression (NMS). Produces the final per-tree polygons.

Requires: infer_gdf with geometry, object_id, tile_path, and detector_score and/or segmenter_score columns

Produces: aggregated infer_gdf with aggregator_score


Classifier

Classifies each detected/segmented tree into categories (e.g. species).

Requires: tiles_path, infer_coco_path

Produces: classification scores and predictions in infer_gdf


Runtime validation

Every component is decorated with @validate_requirements. At runtime, before the component logic executes, the decorator checks that all required state keys and GDF columns are present — and raises a clear error with hints if anything is missing.