| Author | Message | ||
     Scott Parsons (aussiecologist) Member Username: aussiecologist Post Number: 12 Registered: 4-2007 |
Thanks everyone for all of your responses and private emails! Peter, I have contacted Graeme re ANISG. Many Thanks! | ||
| Marion Cuny (marion) New member Username: marion Post Number: 1 Registered: 6-2009 |
Hi Scott, when you want to find a relationship between spectra and other measurements the prefered method is PLS, and the algorithm that enable to take into consideration several Y variables is PLS2 in Unscrambler. Then you have to study the Bcoefficient for each model created. In Unscrambler your settings will be X: spectra, Y: chemical physical measurements you have to weight the Y data if it is not the same scale : Weight : 1/Std you can also use the Uncertainty test that is available in the menu. In this way you will get an overview of the important variables (wavelenghts) that are changing in relation with the chemical and physical info. For the classification: as you are looking at differences PLS2 is also the algorithm you should use. you need to set up a category variable Time with Time1, Time2... then you use the option Split category variable in the Edit menu. you set the PLS2 as X=spectra, Y=[Time1, Time2...] weighting is not needed as the scale is the same for all: values are 0 and 1. same thing with the Uncertainty test and the use of the Bcoefficients to know what are the wavelenght showing the degradation at the different time. For the pre-treatments, my colleague Frank Westad suggested to do an EMSC with background information. One option could be to take the average of the leaves before decomposition and use as a reference spectrum in EMSC. I hope this information will be usefull to you. if you have more question you can contact me directly [email protected] (I am chemometrician for CAMO and so a super user of Unscrambler ;) ) Have a nice day Marion  | ||
| Peter Flinn (peterflinn) New member Username: peterflinn Post Number: 1 Registered: 1-2006 |
Hi Scott, I assume you are an Australian and an ecologist! If so, you should be aware of the Australian Near Infrared Spectroscopy Group (ANISG). I suggest that you contact ANISG Secretary Graeme Batten so that you will be on our distribution list for Australian and other NIR events. Our next conference will be in Adelaide in April 2010, and it would be good if you could attend! Graeme's e-mail address is: [email protected] Regards, Peter Flinn, Chairman, ANISG/ICNIRS | ||
| Scott Parsons (aussiecologist) Member Username: aussiecologist Post Number: 11 Registered: 4-2007 |
Tony, thank you very much for your response! It was both insightful and inspiring. I have obtained reflectance readings for all of my samples, I did not attempt transmission. I have also been using MSC and EMSC (unscrambler) for scatter correction. This, with 1 derivative calculations, has allowed me to produce good pls models of various chemical components of the leaves both in their initial state, and following decomposition. What I want to do now is interpret actual changes in the spectra occurring during decomposition, and also relate these to changes in the chemistry, and hopefully also the environmental conditions in which they broke down during decomposition (soil chemistry, climate). I understand the power of pca to explore this. However, do you think that a pls discriminant analysis might also be of use (thanks private message), particularly in that it would add time as a variable? Would either of these approaches be better suited to relate changes in the spectra during decomposition to actual chemical changes? Would it be possible to use information such as the coefficients of regression from pls-d and run a linear regression of these against coefficients of regression from pls models I have for actual chemical components? I guess this could also act as additional information supporting or not the predictions I make on the leaves using my pls models. Are there other ways to relate the differences between NIR spectra to the actual physical and chemical changes occurring? Thanks again! Scott | ||
| Tony Davies (td) Moderator Username: td Post Number: 191 Registered: 1-2001 |
Hi Scott, This is a challenging but hopefully stimulating proposal! First of all you will have to get the spectroscopy right. Initially, the leaves will have good transmission properties and it might be appropriate to make transmission measurements but as the browning reactions get going absorption will increase until the only measurement you can make is by reflection and this will eventually become a surface measurement. My thought is that you need to alter the samples so that all your measurements are at approximately the same level of absorption. I�m thinking of grinding to reduce the particle size and dilution (sulphur?) but you will think of others. You will still need to get the pre-processing of the spectra right to equalise (as much possible) the effects of scatter. I am a great believer in the power of PCA. I think if you obtain a PCA for each set of samples at each time interval then you should be able to map how these move along your time scale and I hope that this will be easier to interpret than the SIMCA. Many years ago, I was involved in a project to predict the taste quality of chocolate from NIR spectra of the raw or roasted beans. We were able to show some progress in this relationship [Davies et al., Vibr. Spectrosc., 2, 161-172 (1991)] but the funding ended before we were able to attempt to discover the chemical changes that we were utilising. I am not sure if anyone has done this successfully on a complex system but it should be one of the goals of NIR spectroscopy! If you can find a PC which is gaining importance then you could attempt to discover its spectral components. Even longer ago, I worked for Professor R.L.M. �Dick� Synge (of Chromatography fame!) on a project to try to understand the chemistry of decomposition. I remember attending a conference with him and when his friends asked him what he was doing their thoughts were �Only Dick would tackle such a complex problem�! My favourite publication from this era is [Davies, Newby & Synge, JSFA 29, 33-41 (1978)] the only time I have been linked with a Nobel Prize winner! The experimental samples were Cuban cigars donated by a Dr Castro (yes, THE Dr Castro!). I will be very interested to hear your progress. Good luck. Tony | ||
| Scott Parsons (aussiecologist) Junior Member Username: aussiecologist Post Number: 10 Registered: 4-2007 |
Hi, I am doing a study looking at changes in leaves during decomposition. I have NIR spectra of leaves exposed on the soil surface over time i.e. in litterbags. In other words, spectra of the initial leaf properties and those after decomposition. I'm looking at changes in the leaves with more direct ways through chemometric calibrations. However, one idea I've been toying with lately is using a classification approach to discriminate between the spectra, and to view the differences. e.g. to view the regions of the spectra that relate to changes in the leaves during decomposition. Using the Unscrambler I create PCA's for each time step (e.g. initial leaf spectra before decomp, and the spectra after exposure) and run a SIMCA classification. I then use the "discrimination power" plot to view the regions of the spectra with the most discriminatory power between the two models. Firstly, can anyone suggest if this is a good/preferred method for exploring this? Are there better alternatives, maybe with Unscrambler? and... What I want to do next is relate these to actual characteristics of the leaves, possibly by using the information within pls models I have for various components, or through regions of the NIR spectra known to relate to various things. Can anyone suggest a way to do this? Scott |