Move a model to a second instrument: bias and slope before and after

We built a model on instrument A, then asked it about the same unseen-season fruit as measured by a second instrument.

No cleared two-instrument dataset exists yet, so instrument B is simulated: the real unseen-season spectra with a gain of 1.06, an offset of 0.03 and a baseline tilt added by us.

Read every figure here as a demonstration of the method, not as a result for a real pair of instruments.

Bias and slope of predicted against reference, on A and on simulated B.

On A the bias is -0.21. On B it jumps to +1.13 and the typical error goes from 1.06 to 1.53.

Use the buttons above to switch between A, B and B corrected.

A correction fitted on 10, 20 or 40 reference samples measured on B and scored on the rest, 500 random draws.

With 20 samples the typical error drops from 1.53 to 1.08 (middle 80 % of draws: 1.01 to 1.16). The bias goes to +0.00.

Bias is the easy part. The slope stays near 0.87 because the model itself shrinks, on A as well.

The same test with 1st-derivative spectra.

Bias on B is +0.89, smaller than raw (+1.13) but not gone; the gain still scales the derivative. After correction the error is 1.03.

Do not assume a derivative makes the model portable. Check on both instruments.

Not yet: direct and piecewise direct standardization on real spectra from two instruments.

We are waiting on permission to publish a public three-instrument dataset.

Join the list and we write when calibration transfer opens in the Lab.