Soft targets & weights

Soft targets & weights converts hard or probabilistic class labels into model-ready supervision layers: soft class targets, uncertainty, confidence and training weights. UST means Uncertainty-aware Soft Targets.

Minimum inputs

Input	Needed for	Notes
Hard label raster	Hard-to-soft mode	Converted to one-hot class probabilities.
Probability stack	Probabilities-to-UST mode	One band per class, normalized internally. UST = Uncertainty-aware Soft Targets.
Class schema CSV	Recommended	Defines class order, names and optional quality values.
Edge, temporal, source, context inputs	Optional	Increase smoothing/risk and reduce reliability where labels are less certain.

Important controls

Control	Effect
Base smoothing alpha and max alpha	Control how strongly the original label is softened.
Edge/temporal/source/context contributions	Weights for risk terms in the smoothing strength.
Quality prior Q	Global prior reliability of the label source.
Class balancing	Optional inverse or square-root inverse class-frequency weights.

Core calculations

α(x) = clip(α₀ + β_ee(x) + β_tr_t(x) + β_sr_s(x) + β_cr_c(x), 0, α_max)

Smoothing strength: risk terms increase label softening. e is edge risk, r_t temporal risk, r_s source risk and r_c context risk.

p_soft,c(x) = (1 − α(x)) · p_c(x) + α(x) · u_c(x)

Soft target: the original class probability p_c is mixed with a smoothing target u_c. The smoothing target is uniform unless class-context information is supplied.

H(x) = −∑_c p_soft,c(x) · ln(p_soft,c(x)) / ln(C)

Entropy: normalized uncertainty of the soft target over C classes.

U(x) = max(1 − max_c p_soft,c(x), H(x))

Uncertainty: combines low class dominance and high entropy into one 0..1 uncertainty layer.

R(x) = Q_eff(x) · (1 − U(x)) · (1 − 0.5e) · (1 − 0.5r_t) · (1 − 0.5r_s) · (1 − 0.5r_c)

Reliability: separates usable training weight from uncertainty. Temporal risk is 1 − q_t when temporal quality is supplied.

w_c(x) = p_soft,c(x) · R(x) · q_c · b_c

Class weight: combines soft membership, pixel reliability, optional per-class quality q_c and optional balancing factor b_c.

Main outputs

Output	Meaning
Soft target stack	One probability-like supervision band per class.
Uncertainty	Overall uncertainty derived from dominance and entropy.
Training weights	Overall reliability/weight raster.
Class confidence, uncertainty and weights	Optional per-class supervision diagnostics.
Report JSON	Inputs, parameters, source mode and output paths.

Related supervision work

The quality-aware and ageing-aware supervision idea is applied in: Hauser, S., Dachsberger, S., Schmitt, A., & Hinz, S. (2026). Calibrated U-Net with HELIX-Based Label Enrichment for Ageing-Aware Spatio-Temporal Urban Change Detection. ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Accepted / in press; DOI not publicly available yet.