| Author | Message | ||
     david (Unregistered Guest) Unregistered guest |
Hi My name's david and I try to do same quality paramters assessment in barley using NIRs such as protein, beta glucan, hardness and husk%. the software used is winisi but I need if it's possible calibrations for beta glucan, hardness and husk% which could give me a good results and a significant correlation. Regards David | ||
| david (Unregistered Guest) Unregistered guest |
Hi My name's david and I try to do same quality paramters assessment in barley using NIRs such as protein, beta glucan, hardness and husk%. the software used is winisi but I need if it's possible calibrations for beta glucan, hardness and husk% which could give me a good results and a significant correlation. Regards David | ||
| Dafna |
I have been developping calibrations for grains (wheat, barley, corn) with poor correlations for fat and fiber. Actually, I would like to know what could be the possible reason for the low RSQs. Is it because of the very low variability of these parameters? Does anyone have a good calibration for these parameters in grains? | ||
| David W. Hopkins (Dhopkins) |
Dafna, There are many different instruments that can measure fat and fiber in wheat, barley and corn. However, there are many reasons why you may not be obtaining as high a R-squared value as you would like. The RSq is perhaps the worst statistic to characterize a calibration, because it is so sensitive to the range of the analyte, and you have not mentioned the range you have available. It is particularly difficult to obtain a range in fiber content, so you may have to start out with a preliminary screening calibration and then find new samples to extend the range. Are you measuring ground or whole grain samples? What methods are you using for the reference determinations of fat and fiber, and what is the SD of repeats of the same sample that your laboratory is able to achieve? Then, what is the comparable performance of your NIR measurement? Regards, Dave Hopkins | ||
| fernando morgado |
Dafna : First is necesary to know if your samples are grinding or not, wich region you are using for measure, and if your measurent are in transmision or reflectance mode. Other thing is if your samples are moved during the scaning, because this type of samples are no homogeneus. If you obtain good calibrations for other parameters probably you have problems with database. Wich is your database range for both parameters, and how many samples do you have are important factors to considerate. You obtain bad correlation after eliminate outliers or before it? You are making spectral preprocesing ( derivates, mean center, etc)? Sometime only one sample with problems give very bad general stadistics, and when you remove this sample the stadistic is better. You check spectral residual, leverage, and score analisys for your samples ? Is dificult indicate wich can be the problem without know the adquisition specifications, database performance or look the spectra. best regards Fernando Morgado Chile | ||
| dafna |
Than you both, David and Fernando! I have been working with a NIRSystems 5000 instrument (reflectance mode) and WinISI software, with two different calibration files for each grain - ground and whole. My reference methods for fat and fiber are ethereal extraction (Soxtec/Foss Tecator) and extraction (Fibertec/Foss Tecator), respectively. The range of my wheat samples, for instance, is 2.0-3.1 for fiber and 0.96-1.77 for fat. I do not believe the reference data is my problem since our lab performance is controlled by ring-tests that confirm the consistency of our results. I should though consider rechecking it. What do you consider being a good number of samples for a calibration set? And, Douglas, could you please tell me what do you mean by "preliminary screening calibration"? | ||
| Peter Tillmann (Tillmann) |
Dear Dafna, there is almost no variability in grains with respect to fiber and fat (at least none compared to protein and starch). If you find variability worth for calibration efforts it is fiber in barley (own experience) and maybe fat in corn (no own experience). If you compare your range to the precision of the fiber and fat determination you will see this clearly. Range of fiber in wheat: 2-3% Repeatability for fiber determination: 0.14% Thus the standard deviation of your results is 0.16% (rule of thumb: 1%/6) and equal to your precision. You need no analysis to get the same results, if you take the mean (as all analytical laboratories will do for fiber in wheat in routine analysis). For fat in wheat the story is the same: range: 0.8% repeatability: 0.08% SD 0.64% = approx. range This has nothing to do with instrument manufacturer or similar. Peter Tillmann | ||
| Fernando Morgado |
Dafna : I don´t know wich module you are using, spining ring or transport. If you are using transport the better cup for ground samples is quarter cup, and for whole sample will be better a natural product sample cell or similar ( more big medition area). If you are using spining ring this is not good for whole samples. The problem can be : 1.- Instrument I supose your instrument is O.K, repeatibility test RMS near to 20, and first data point of polyestireno in 1143 or data point 80, and correct check cell result. 2.- Spectra : Check the spectra collected are correct, you can look it in spectra option, maybe you have some spectra very diferent or bad. 3.- Database : Check you indicate the correct laboratory data in samples management. Sometimes is possible make a mistake when you type the laboratory data. 2.- Calibration I indicate some standard configuration for develope calibration using WINISI in food a.- mean center on b.- scatter correction on c.- derivate 1,5,5,1 d.- Number of pass elimination 3 e.- Limit H : 3 f.- Limit T : 2.5 g.- Limit x : Increase to 10. h.- Number of factors : use the default. i.- Indicate don´t stop for analize result. J.- range : intent the follows range 1100-2498,8 ( standard for food) 1100,2498,4 1100,2498,2 3.- After run the calibration check the number of factors calculate for your model ( you can see it in calibration management). The number of factors is a important data. When you don´t have good correlation the number of factor are very low ex. 1 or 0. If your data base have few samples you can obtain 2 or 3, but the number depend of the parameters. For example for moisture using 100 samples the factors can be 4 or 5. 4.- After run the calibration go to monitor calibration and look the result and correlation graph. 5.- About the number of samples depend of parameter but in my experience always you need 60 to 100 samples, sometimes more.( considering you have a complete range) Best Regards Fernando Morgado | ||
| David W. Hopkins (Dhopkins) |
Dafna, I agree with the points made by Fernando Morgado and Peter Tillmann. I would say that Peter is very fortunate if his lab has a 0.14% SD on fiber. My experience with TDF was that we obtained more like 0.3%(on bran fraction from mills, not whole grain), and I was happy with a SEP of 0.5%. Then, if the range of fiber in whole wheat is 2%, the ratio of the range to SEP is only 4, which is useful but not very great. I like to see this ratio at 6 or more for a good calibration. There are only 2 ways to improve such a situation, find ways to decrease the SEP, or obtain samples at the high and low ends of the range to extend the range. We usually try to do both. The best way to find the true lab SD is to submit blind duplicates, and determine the total variability in sample preparation, sampling, and determination. The calculate the SDD (standard deviation of differences) on the pairs of results. Usually 5 to 10 blind duplicates are sufficient to see the situation. Don't just assume that you have no problem with the lab results. When the ratio of the range to the SEP is less than 4, the calibration may not have a lot of predictive value, but it is still capable of identifying samples at the high and low ends of the range. This is what I would call a screening calibration, it is useful for screening samples for further calibration development. Regards, Dave |