RPD explained: is your NIR model good enough?

A plain explanation of RPD and SEP in NIR calibration, with the rule-of-thumb thresholds for QC and a quick check you can run on your own model results.

The short answerRPD is the standard deviation of the reference values divided by SEP. Rule of thumb: 3 or more is acceptable for QC, 5 or more is excellent, under 2 is poor. Always report it with RMSEP, never R-squared alone.

Key numbers

RPD 3+acceptable for QC
Course glossary
RPD 5+excellent
Course glossary
RPD under 2poor
Course glossary

This is a sample post that shows the template. The real rewritten posts from the blog project drop into this folder with the same front matter.

Why RPD exists

SEP alone has units, so a 0.3 SEP means different things for protein and for moisture. Dividing the spread of the reference data by SEP gives a scale-free score.

How to compute it

  1. Hold out an independent test set.
  2. Take the SD of its reference values.
  3. Divide by the SEP of the predictions.

Quick check

Reference SD is 1.2 and SEP is 0.4. What is RPD, and is it good?

Ask Spectra

Go deeper in the course

L23
Introduction to Calibration
Week 6 · 18 min
See in curriculum
L24
Calibration Validation Techniques
Week 6 · 19 min
See in curriculum

Enroll — $99 lifetime  14-day refund

Sources

  • SpectroScience course glossary: RPD, SEP, RMSEP entries

Questions

Is a high R-squared enough?

No. R-squared can look good on a weak model. Cite RMSEP and RPD from an independent test set alongside it.