| Author | Message | ||
     Howard Mark (hlmark) Senior Member Username: hlmark Post Number: 361 Registered: 9-2001 |
Magda - Prusisan blue is insoluble in water, but perhaps there's another solvent that might dissolve it. You stated that you measured the spectrum of "pure" Prussian Blue; did you mean as a solid powder or as a dispersion of pure Prussian Blue in water (or some other solvent)? Another approach might be based on one of the uses found on the Wikipedia page: http://en.wikipedia.org/wiki/Prussian_blue They mention several chemical reactions involving Prussian blue. For example, one is a reaction involving phenol. Thus there is a possibility of using a solution where excess phenol is present, and then the result of the reaction would depend on the amount of Prussian Blue. Possibly similar use could be made of one of the other reactions discussed. \o/ /_\ | ||
| Magdalena Sut (vickym) New member Username: vickym Post Number: 3 Registered: 11-2010 |
Howard-the problem is that Prussian Blue is not water soluble, what kind of extract would you suggest? The other thing is, when I scan pure Prussian Blue I can see very characteristic spectrum, with the peaks at 1900nm,2030nm, 2090nm, 2145nm, 2370nm. Can it be that the last peak is acctually the CN bond?I can see it in the samples with higher contamination from cyanide (17000-10100mg/kg) at 2350nm, but with the lower concentrations it disapear. We first started the experiment with mixing the prussian blue with the pure quarz, the calibration worked then. But when it comes to real soil samples from the MGP, where we know for sure that we have prussian blue there (which is not soluble) why we cant see it? we even have the samples that are literally blue..why is that so? is the soil matrix to complicated? is there an overlap of other overtones? | ||
| Howard Mark (hlmark) Senior Member Username: hlmark Post Number: 360 Registered: 9-2001 |
Magda - I have to agree that your system probably does not lend itself to standard NIR analysis. On the other hand, Prussian Blue is a very strong absorber, so that the lower limits of measurement we commonly encounter may not apply. BUT: Prussian Blue is a very strong absorber - - - in the VISIBLE region of the spectrum, so maybe you should try making measurements there, instead. And, although it flies in the face of "normal" usage of NIR and of spectroscopic analysis as it's commonly applied these days, it may be helpful to apply some simple chemical methodology and, for example, try extracting the Prussian Blue from a weighed amount of sample and measuring that in solution, regardless of which spectral region you use. Howard \o/ /_\ | ||
| Magdalena Sut (vickym) New member Username: vickym Post Number: 2 Registered: 11-2010 |
Dear all, I am working with samples coming from Former Manufactured Gas Plant, so the iron that is present there is mainly in form of iron-cyanide complexes (mainly Prussian Blue). My second task is to determine CN bond with NIR, but about it maybe I will write later. Comin back to iron, the concentration range is from 7-131 g/kg, in my training set 7-131 g/kg and in validation set 14-63 g/kg. When I applied Full cross validation to the whole data set (train+validation), the RMSECV=26 with factor no=2. In training set, plot of predicted values vs actual has RMSEC=17, and cross validation vs actual=30. The wavelenght rage of my data is 1595-2396nm. S.Golay does first and second derivative and smoothing. The optical resolution of my spectrometer is 11nm and it covers the spectral range of 1600-2400nm. I have to admit that I am quite new in this field and its my first experiment with NIR and chemometrics  Magda | ||
| Gabi Levin (gabiruth) Senior Member Username: gabiruth Post Number: 45 Registered: 5-2009 |
Hi Pierre, Thanks for the information - it substantiates what I was trying to say - first understand what is the chemical formula, then the concentration. Both these factors go against any reliable NIR measurement of iron in soil, because as you said, it goes with organic matter, that I assume you mean that all iron is organically bound - I am not sure about- but let us assume so, the problem is that not all organic is bound to iron, much of it has nothing to do with iron. If we could really assume that the 10 ppm of iron is all bound to organic matter as a salt of organic acids, the question is then - could that little iron affect the spectrum sufficiently to be measurable - I strongly doubt it and this is probably the cause of the poor prediction capability. To be absolutely honest, I have seen people try to use NIR to measure almost everything under the sun, and many times only because it is so "EASY", just place samples, collect spectra, run PLS and abra cadabra - here you have a calibration and life becomes so beautiful. This is done so many times without any serious attempt to understand why it works when it does, nor why not when it doesn't. NIR is not a magic tool, it works when the phenomenon we try to measure conforms to certain rules that make it possible to do so, but these rules are a whole different story. Gabi | ||
| Pierre Dardenne (dardenne) Senior Member Username: dardenne Post Number: 52 Registered: 3-2002 |
Hi Magdalena, Gabi's calibration scheme is nice, but iron concentration in soils in less than 10 ppm and then NIR is unable to see this level. There is a correlation with organic matter (OM). If you get a correlation for iron, it is the one likely with OM. OM is predictable by NIRS never iron. This is valid for many other elements in soils. Pierre | ||
| Gabi Levin (gabiruth) Senior Member Username: gabiruth Post Number: 44 Registered: 5-2009 |
Dear all First thing we need to ask ourselves is - what is the chemical formula in which the iron is present, and if there is a good reason to expect that this formula will give rise to significant spectral features that can be correlated to iron. The next thing we need to find out is what is the concentration range, and estimate if this range is likely to have a sufficient finger print in the spectrum. The next thing is to find if the uncertainty (I prefer this term to accuracy) in the reference values is sufficiently small as compared to the range. The uncertainty shall not exceed 5% of the range of the iron. If it exceeds 5% of the range, the potential for getting good regresssion is greatly diminished. Then we need to ask ourselves if 30 samples are really sufficient to represent all the variability in the iron and other possible parameters that may vary simultaneously with the iron. It is possible that the sample set is too small and the initial regression is not really robust enough. My first test will be to run a regression with all 40 samples, using cross validation as a first attempt to evaluate the prediction capability. If the cross validation for all 40 samples still looks good, the next step should be to add 30 to 40 more samples, increasing the range if possible, and running at first all the set with cross validation and again evaluating the quality of the cross validation. If it is still good, then it makes sense to try to use a separate validation set, by predicting some 20 samples from the model and calculating the validation error. Trying to place much emphasis on various spectral treatment will not solve a problem if the source of the problem is residing with one of the reasons I mentioned above. Usually 1st derivative with 9 or 11 smoothing elements is sufficient if the spectrum contains real features that are uniquely related to the iron concentration. I have seen too many times that initial small sample set regressions look good, but once faced with suficiently large data set will fall apart because there is no real reason for it to work. Gabi | ||
| David W. Hopkins (dhopkins) Senior Member Username: dhopkins Post Number: 177 Registered: 10-2002 |
Hi Magdalena, Welcome to the Discussion Group. I'm sure you already know that you have picked a very challenging application. We need a bit more information to help you. What are the ranges of the Fe concentrations in your training and validation sample sets? What are the units? What is the precision of your reference method, as a standard deviation of replicates? Have you compared the accuracy you can achieve using PCR? When you say you are using "S.Golay", I assume you are using the smoothing function. Have you tried first or second derivatives, and then SNV? What is the wavelength interval for your data points? What resolution are you using in your spectrometer? These details help to optimize the S-G parameters. Best wishes, Dave | ||
| Magdalena Sut (vickym) New member Username: vickym Post Number: 1 Registered: 11-2010 |
Hello, I am working with FP NIR and I am trying to build up a calibration model that is going to detect iron concentrations in soil samples. As a calibration software I use Phazir MG. I have 40 samples. 30 I used as a samples for model delevopment and 10 I left for testing. I preprocess my spectra using S.Golay and SNV, than I use PLS. When I create the model the RMSEC is 5 with factors no 6. I dont know why, when I use the test samples to validate the model the correlation coefficient is very low. How can I improve my model? |