For quantitative purposes, hyperspectral imaging seems to lead to a bigger prediction error when comparing its prediction ability to traditional NIR instruments. Despite its advantages and potential, this seems to me a relevant issue for a full implementation of this technology.
Which factors are in your opinion the most important ones in leading to lower performance? There might be the effect of sample position, illumination condition, the difference which is obtained when looking at single pixel level or averaging areas (especially when applying the calibrations!), the fact that the sample is moving (in a push-broom system, for example), and further error is obtained when looking at single objects instead of average batches.
Thus, are there actual benefits of using HSI over traditional instrumentation to study the composition of food and agricultural products?
The only approach is to get reference for an object (kernel, bean, etc.) to make reference measurements. Now, on this aspect, one can choose to build a model using all the pixels of the objects and attributing the reference measurement (1 sample) to those spectra (thousand of samples); the other approach can be to average the object and just use 1 average spectrum.
Then, once the calibration is developed, there is the same dilemma, i.e. to apply on single pixels or on averages. But even considering averages only, so working like a traditional NIR instrument, there seem to be in the literature quite a big difference in the prediction ability.. Isn't it?
There are many factors affecting the accuracy of VNIR hyperspectral imaging.
First is your samples and the way you do the reference analysis. Let's suppose you are working on kernels to measure for example N content. How many kernels do you use for measuring N content using your ref method? If 5 kernels, then you should use the average spectra of all the 5 kernels to correlate to the reference N (There are more approaches but I think this is the best for linear models).
Second, your camera properties. I mean resolution and spectral range of your camera? Smaller resolution and wider range are more favourable.
Third, outlier detection and data pre-processing are so important. Dimension reduction and/ band selection also improve the accuracy of the models.
Fourth, the model you use for calibration. Is that a linear model or a non-linear? Most of the times linear models fit well to your data but sometimes you need to use nonlinear models.
The number of the samples you are using for calibration is also important more specifically when you are using machine learning models such as ANN.
Once you develop the model, then you can use any pixel for prediction in new samples. Even you can use this system Real time.
Hope I could add anything to what you already know.
- Badges
- Science topic
- Similar topics
- Geoscience
- Atmospheric Sciences
- Atmosphere