Context features

Context features describe the neighbourhood of each label pixel: edges, local diversity, probability confidence, local class support, purity and optional class-pair relations.

Minimum inputs

Input	Needed for	Notes
Hard label raster	Edge risk, diversity, hard-label class context	Uses the selected NoData/background value.
Probability stack	Entropy, margin, soft class context	One band per class; stack is normalized internally.
Class schema CSV	Optional	Gives readable band names and class order.

Important controls

Control	Effect
Radius and window shape	Define the local neighbourhood `N_R(x)`.
Feature set	Selects edge, diversity, entropy and margin.
Write class/local/pairwise context	Creates optional multi-band context stacks.
Include self-pairs / max pairwise bands	Controls class-pair stack size.

Core calculations

edge(x) = 1 if any 4-neighbour has a different valid class; otherwise 0

edge-risk_R(x) = dilate(edge, R)

Edge risk: marks class boundaries and their neighbourhood. It helps identify pixels where a crisp class label is spatially fragile.

diversity_R(x) = |{z∈N_R(x) : y(z)≠y(x)}| / |{z∈N_R(x) : y(z) valid}|

Diversity: fraction of valid neighbours that differ from the centre class. High diversity means mixed local context.

H(x) = −∑_c p_c(x) · ln(p_c(x)) / ln(C)

Entropy: normalized probability uncertainty. It requires a probability stack and is 0 for a confident single class.

margin(x) = p₍₁₎(x) − p₍₂₎(x)

Margin: distance between the strongest and second-strongest class probability. Low margin indicates ambiguity.

support_c,R(x) = mean_{z∈N_R(x)} p_c(z)

support_c,R(x) ← support_c,R(x) / ∑_j support_j,R(x)

Class support: local average support for class c, normalized over classes where support exists.

purity_R(x) = max_c support_c,R(x)

Local purity: dominance of the strongest class in the neighbourhood.

pair_i,j,R(x) = p_i(x) · support_j,R(x)

Pairwise context: describes how centre-class support i interacts with neighbourhood support j. Self-pairs can be included or excluded.

Main outputs

Output	Meaning
Edge risk	Byte 0/100 raster, NoData 255.
Diversity, entropy, margin	Float32 0..1 rasters, NoData -9999.
Class support / local purity / pairwise context	Optional multi-band context stacks for later modelling or supervision.
Report JSON	Parameters, skipped outputs, band counts and paths.

Entropy and margin require a probability stack. Skipped outputs are reported and leave empty result paths.