NIR Calibration, Validation and Transfer
How to build, validate and maintain an NIR calibration: sample selection, preprocessing, PLS modeling, independent validation and calibration transfer.
Key numbers
Illustrative spectrum for teaching, not measured data.
Scan the sample to light up the bands, then pick a constituent.
Quick check
Matching lesson: L23Build
Start with at least 50-100 samples that span every source of variation: raw materials, process, season and storage. Pick the number of PLS factors by watching RMSECV. Stop adding factors when it stops decreasing.
Preprocess
SNV corrects scatter one spectrum at a time and is the default for most ground samples. MSC uses the mean spectrum of the calibration set. Savitzky-Golay derivatives remove baseline drift.
Validate
RMSEC is always optimistic. RMSECV is estimated inside the calibration set. RMSEP, measured on an independent test set, is the one to trust. RMSEP clearly above RMSECV means overfitting. Never cite R-squared without RMSEP and RPD on the same data.
Transfer and maintain
To move a model to a new instrument, measure a standard set of 10-20 diverse samples on both and correct the difference. Afterwards, check a few samples against the reference regularly. A constant bias gets a slope and bias adjustment; a bias trend means recalibrate.
Go deeper in the course
Enroll — $99 lifetime 14-day refund
Read more on the blog
- How to Transfer NIR Calibration Between Instruments Effectively
- NIR Calibration for Better QC: Reducing Drift and Improving Accuracy
- PLS Regression for NIR: Step-by-Step Guide for Food and Feed Calibration
- Why Does R² Lie? Understanding RPD, SEP, and RMSEP in NIR Accuracy
- NIR Spectral Preprocessing: SNV, MSC, and Derivatives — When to Use Each
- The Systematic Approach to Diagnosing NIR Calibration Problems
Links open the current site. Posts move here later.
Related
Questions
What is a good RPD?
As a rule of thumb, an RPD of 3 or more is acceptable for quality control, 5 or more is excellent and under 2 is poor. RPD is the standard deviation of the reference values divided by SEP.
SNV or MSC?
Both correct scatter from particle size and path length. SNV works on each spectrum alone and is the usual default for ground samples. MSC refers to the mean calibration spectrum.
How do I transfer a calibration to a new instrument?
Measure 10-20 diverse samples on both the master and the new instrument, then correct the differences. Copying a model across unchanged often fails because of small optical differences.