Support, confidence and consistency

Every pixel of a super-resolved image mixes two things. The first is what the 10 m measurement determines. The second is what the model infers from learned priors. synapse-sr keeps the two apart and reports them.

Decomposition

import synapse_sr
result = synapse_sr.super_resolve("scene.tif")

result.x_base   # determined by the Sentinel-2 observation
result.prior    # contributed by the learned prior; invisible to the sensor

x_base is a deterministic, regularised inversion of the sensor model. prior lies in the null space of the forward operator: re-observing the image with the Sentinel-2 model gives the same measurement with or without it.

Support classes

result.support measures the prior's contribution against each band's sensor noise level tau_b:

Class	Value	Rule	Reading
HIGH	2	max over bands of |prior| / tau_b < 3	essentially observation-determined
MEDIUM	1	between 3 and 10	the prior adds noticeable structure
LOW	0	above 10, or invalid input	prior-dominated: treat as inferred

The thresholds are heuristic and are reported in result.metadata["support_classes"].

Calibrated uncertainty

import synapse_sr
result = synapse_sr.super_resolve("scene.tif")

err = result.uncertainty()        # (4, 5H, 5W) expected absolute error, reflectance
half = result.interval(0.9)       # reference within image +/- half with probability 0.9

uncertainty() comes from an error model shipped with the checkpoint. It regresses log |error| on the learned error scale, the prior's magnitude relative to sensor noise, local edge strength and variance, brightness, NDVI and band, against a 2 m HR reference (GeoSR development split). interval() scales it with split-conformal quantiles. Measured on development patches not used for fitting or calibration, for Pro v2 (Pro v1: 82 / 91 / 96 %):

Level	Measured coverage
80 %	82 %
90 %	91 %
95 %	96 %

The error map also ranks error. Keeping the 50 % of pixels it marks as most reliable roughly halves the mean absolute error, and its rank correlation with the actual error is 0.58.

result.confidence is the network's raw error-scale output. Use uncertainty() or interval() for decisions.

Consistency

result.consistency re-observes the output with the Sentinel-2 forward model and compares it with the input, per band, in units of that band's noise level:

{"B04": 5.7, "B03": 3.5, "B02": 3.0, "B08": 3.2}

The physics baseline deliberately fits the measurement to about 4 noise units, not 1: a real scene also carries forward-model error (point-spread function, registration), and fitting it to sensor noise alone turns that error into false fine detail. Because one regularisation weight is shared by all bands, individual bands can sit a few units above or below 4. Values far beyond that (above 5 x discrepancy) trigger a warning: check the band order, the radiometric offset and that the input is L2A on the 10 m grid. super_resolve(..., discrepancy=1) fits to sensor noise alone. The learned detail itself is invisible to this measurement by construction (it lies in the null space). The measurement is taken on the assembled mosaic, so tiling seams would show up here.