Calibration steps summary

Ganya's picture

Hi All,

I am about to create calibarations on NIR and I have collected data from the reference method which is wet chemistry analysis in my case.

How do I go about creating calibrations?

I would really apreciate if one could send me the steps to follow.

Many thanks. I am looking forward to your responses.

shileyda's picture

Hello Ganya,

You are skipping over a most important (or the most important) part of creating a calibration...the samples and reference data!

Of course there are many excellent textbooks on this entire subject and I highly recommend reading several to better understand the entire process. Very basically here is what I recommend.

1. Identify the lab and method that you would like to use and test the capabilities of the lab. This is done by splitting a set of at least 6 samples into three portions, resulting in 18 total samples. Randomize the sample ID for these samples then submit these to the lab. Calculate the reproducibility of the method. If this meets your objective continue with the lab, if not, find a new lab. Another critical question at this time is regarding the sample moisture and chemical stability. Are these subject to change or degradation from when they were collected to when they are tested in the lab? If so what steps can be taken to assure that they do not change?

2. Obtain samples that contain maximum variation in spectral characteristics and in chemical composition. This is the most challenging part, and often the part of the process most ignored. Do this poorly and your model will not represent your process or population of samples. Also unless you make some intelligent decisions about your sample set selection you will spend more money than you need to and get an overall poorer calibration result due to the lack of diversity contained in your calibration set. Various tools exist to select appropriate samples by their spectral characteristics, such as nearest neighbor. I highly recommend that you use such a tool.

3. Decide on your sample presentation and sampling accessory. Some experimentation is needed to decide on how best to collect the spectra according to your desired workflow. Will the sample be in solid, liquid or gas form? Will the sample be non-homogeneous that will require a very large sample? Will it be ground? If so what screen size and grinder should be chosen? What speed of analysis is needed? Where will the spectra be collected? Will this be in the lab, at the production line, in-line or some field measurement? Each of these also has a set of considerations for optimal spectral collection.

4. Choose a modeling software package - this is like asking which is the best car...I'm not going there... consider whether you want every possible feature or whether basic algorithms and simpler operation are sufficient.

5. Divide your available data into separate calibration and test sets. Again many approaches here, I prefer to sort my samples by the constituent and assign every 5th sample to a test set. How many go into the test set depends on how much data you have availble. I think 20% is a bare minimum for the test set though.

6. this is the fun part.  Be creative but be sure to evaluate various spectral pretreatments and also be sure to evaluate using only portions of the spectrum. Compare SECV, RSQ and the number of factors in your models. Consider removal of outliers due to lack of fit in either spectral or concentration residual, or better still, understand why these samples varied and either add more similar type of samples to better represent them or retest in the lab.

7. Validation - Once you have a chosen you top candidates from the last step, now it is time to compare the independent test set results for SEP, RSQ, RPD and spectral residual measurement. Does the performance of the test set resemble he calibration result? If so, then deploy the model, if not, then ask why that happened and iterate.

8. Monitor - Develop a plan to monitor the performance of your models. This should include submission of samples to the lab method that had both good spectral deviation and samples that were poorly represented. Calculate SEP on ongoing basis and if this begins to rise it may indicate the need for a model revision due to some changes in the spectral characteristics of the sample population.

9. Review and revision - Part of your monitoring plan should include the frequency of model review and what conditions will prompt a model revision. But you should expect to need a revision no matter how carefully you selected your initial samples for the model.



Ganya's picture

Hi Dan,

Thanks a million.

Your response is loud and clear, I appreciate.

ianm's picture

There are two scientific society methods describing the calibration process:

AACC 39-00 and AOCS Am1-09

Posted by Ian Michael on behalf of:

Charlie H
Dr. Charles R. Hurburgh, Jr.
Professor, Agricultural and Biosystems Engineering Professor in Charge, Iowa Grain Quality Initiative
3167 NSRIC
Iowa State University, Ames. Iowa  50011

ianm's picture

Here are the basic steps (see uploaded file), Ganya. Happy to answer any questions.


Posted by Ian Michael on behalf of Phil Williams

Uploaded Files: 
Ganya's picture

Hi Phil,

Thank you very much.

I did identify the most suitable refence method and analysed my samples. soon I will be scanning my samples on the NIR.

I appreciate your assistance.


hlmark's picture

Ganya - Dan, Charles and Ian are giving you excellent advice, in fact, that's a summary of everything you will read on the subject of calibration, in a nutshell. One point that may be missing, or at least not obvious in Dan's Monitoring, Review and Revision discussion, is the need to institute a post-calibration QC procedure for your instrumental readings; in fact I think that is what Dan was getting at. My rule-of-thumb recommendation for that is to follow Dan's recommendations for what to do. Additionally, here's some advice for when to do it: as long as there is no indication of a change in the performance of the instrument, perform the QC procedure every day for at least a week, then every week for at least a month, then at least once a month thereafter. This will give both yourself and everyone else confidence in both the short- and long-term stability of the instrument and the readings. If there's a change in the readings that requires a change to the model, then roll back the QC procedure and restart it once you've settled on the new model to use.