HELIX Labelling Framework – overview

HELIX prepares categorical geospatial labels and label-derived supervision layers on a common reference raster grid. The grid may come from imagery, a classified raster, a DEM-derived product, an analysis grid or any other georeferenced raster that defines the desired output geometry.

The workflow is modular: check inputs, reconstruct labels spatially, align label timing, add cyclic time features, derive context information, and create soft targets with uncertainty and weights.

Main engines

#	Engine	Main purpose	Typical handover
1	Preflight & class schema	Checks CRS/grid consistency, class fields, class schemes and remapping tables.	Class schema CSV/JSON and QA report.
2	Spatial reconstruction	Brings vector and raster labels onto one reference grid.	Hard-label raster, one-hot/support stacks, coverage/agreement rasters.
3	Temporal reconciliation	Matches target dates to label snapshots or validity windows and can materialise matched label sources per date.	Date axis, match table, temporal quality and reconciled label-source index/folder.
4	Helical features	Encodes dates as cyclic seasonal features and optional interaction rasters.	Time-feature table and optional raster stacks.
5	Context features	Derives neighbourhood, boundary, class-support and probability-context features.	Context rasters for inspection or later supervision.
6	Soft targets & weights	Creates soft class targets, uncertainty, confidence and training weights.	Model-ready supervision rasters.

How the engines interlock

HELIX separates decisions deliberately. Temporal Reconciliation decides which label state belongs to which target date and writes temporal reliability. Spatial Reconstruction turns the selected vector/raster source into a grid-aligned label product. Context Features describe neighbourhood ambiguity, and Soft Targets & Weights combine class evidence, temporal quality and spatial/context risk into uncertainty-aware supervision. Helical Features can be added whenever date-aware or seasonal model inputs are needed.

Model Builder and scripts

Manual workflows can use the normal output folder. For automation, keep the folder as TEMPORARY_OUTPUT and connect the explicit advanced output destinations, for example hard-label, probability stack, temporal match CSV or report JSON.

Reference conventions

The reference raster defines CRS, extent, pixel size, width and height.
Vector class fields are explicit; HELIX does not silently burn feature IDs, the first attribute or a fallback class value.
Class-schema CSV files can map old/source values to new target IDs and readable class names.
For continuous 0..1 products, value 0 is valid; Float32 outputs therefore use -9999 as NoData. Byte one-hot outputs use 255 as NoData.

Academic background

The general idea of treating label preparation as an explicit optimisation and quality-control problem follows: Hauser, S., Augner, L., & Schmitt, A. (2025). Perfect Labelling: A Review and Outlook of Label Optimization Techniques in Dynamic Earth Observation. Remote Sensing, 17(7), 1246. doi: 10.3390/rs17071246.

The broader methodological context and environmental-monitoring workflow are developed in: Hauser, S. (2025). Automated Feature & Label Refinement in the Context of Environmental Monitoring by Multi-Modal Satellite Data. Dissertation, Karlsruher Institut für Technologie. doi: 10.5445/IR/1000185980.

Export/report and batch recipe remain available under Utilities in the Processing toolbox for handover or advanced repeatable runs.