Spatial reconstruction

Spatial reconstruction aligns vector and raster label sources to the reference grid and can fuse several sources into a hard-label raster plus optional support and QA layers.

Either-or rule: use either a main/reference vector or a main/reference raster. If both should be combined, put one in the main slot and the other under additional sources.

Minimum inputs

Input	Required?	Notes
Reference/grid raster	Yes	Defines output geometry.
Main/reference vector	Required for vector workflow	Needs an explicit reference class field.
Main/reference raster	Required for raster workflow	Can be remapped through a class schema.
Additional vector/raster sources	Optional	Fused after alignment to the same grid.
Class schema CSV/JSON	Optional	Maps source values to target IDs and names.

Important controls

Control	Effect
Vector pixel assignment	Pixel centre, all touched, area majority or minimum-overlap assignment.
Fusion rule	First valid, last valid, majority, weighted support, consensus only or primary-if-valid-else-majority.
Minimum source coverage	Required fraction of selected sources that must provide a valid label.
Treat 0 as class	Use for binary or zero-based class rasters where 0 is a real class.

Core calculations

coverage(x) = n_valid(x) / n_sources

Source coverage: fraction of selected label sources that provide a valid class at pixel x. It is used to reject pixels with too little support.

p_c(x) = #{i : y_i(x) = c} / n_valid(x)

Class support: empirical support for class c among valid sources. For weighted support, counts are replaced by source weights.

c^*(x) = arg max_c #{i : y_i(x) = c}

Majority class: selected class with maximum source support; ties are resolved by the selected fusion rule and source order.

agreement(x) = ∑_i w_i · 1[y_i(x) = c^*(x)] / ∑_i w_i

Agreement: weighted strength behind the selected class. It is high when most reliable sources select the same class.

overlap_g,p = area(g ∩ p) / area(p)

Vector overlap: fraction of target pixel p covered by geometry g. Area-majority assignment picks the class with largest overlap; the minimum-overlap option rejects weak intersections.

purity_R=1(x) = |{z ∈ N₁(x) : y(z) = y(x)}| / |{z ∈ N₁(x) : y(z) valid}|

Immediate purity: checks whether the 1-pixel neighbourhood is homogeneous. Low purity indicates edges or mixed areas.

Main outputs

Output	Meaning
Hard label raster	Single-band target class ID raster.
One-hot stack	One band per class: 1 for membership, 0 for valid absence.
Class probability/support stack	Per-class source support `p_c(x)` in 0..1.
Coverage and agreement	QA layers describing source availability and source agreement.
Class schema and report	Reusable mapping and reproducibility metadata.