- This topic has 5 replies, 2 voices, and was last updated 8 years, 6 months ago by
.
-
Posts
-
I’m a new user, so very inexperienced with Scio. Apologize in advance if any of this is dumb.
I’m trying to get SCIO to detect different samples of rice. I used “my mini applets” feature on the Scio App to create the “rice” applet.
I scanned 4 different kinds of rice using the small items container (I fill the container full with rice) and created the spectro scans.
However, i noticed that the scan varied a lot based on the level of the rice that i fill the container with. The difference in the spectro scan for the same rice but different container levels is SO MUCH more significant than the difference among the 4 samples of rice. Sometimes, the different placement of the rice grains could cause variability in the scan that are more signficant than the actual scanning differences of the rice samples.
I have two questions:
1) any technique to enforce consistency of the scan? for example, should i make the rice into powder form and use a consistent level (for example, use 50g of rice powder)? or maybe turn into liquid form?
2) How different should the spectro scan be for different types of rice? They all look fairly similar to me, and inconsistencies in testing could swamp the actual difference of the samples themselves! The rice actually look quiet different to the naked eye. The composition of the rice are similar, but differences in certain things such as carbohydrate and protein among the rice types far outstrips the 1% sensitivity threshold.
Thank you for your help.
Hi Guowei,
The Consumer app’s workshop applet allows you to work with the spectrometer by collecting material data and building analysis applets for your own materials. However, no customized pre-proccessing methods and algorithms are provided.
SCiO Lab offers various filtering and analysis tools which should be used in order to develop a successful and robust model.
According to our records, you have a developer license, hence using SCiO lab tools is highly recommended.
As for the questions:
1. Consistency is terms of measurement conditions is an important key factor for a successful data collection.
As you have suggested, using consistent amount of rice samples (or more precisely consistent layer depth) is recommended, so that the optical path of the reflected signal from each one of the rice samples will be identical.
2. Generally speaking, the differences between the spectrum curves should be visually inspected when reviewing the reflectance spectrum. Different pre-processing methods should be used: “Normalized” is useful mainly for classification models. Sometimes use of both “processed” and “normalized” is needed when there is a variance in the optical path of the samples.
Choosing the right pre-processing method is a trial and error process, it is recommended to try each one of the methods and evaluate the model’s performance accordingly.
I hope it was helpful.
Let me know if you have further questions.
Ayelet,
The Consumer Physics Team
Thank you Ayelet.
Which pre-processing method is best for powder form? I plan to crush the rice grains into fine powder and use the small items container to collect data. This would allow me to control the consistency of the measurement conditions.
Also, i saw that i need at least 40 samples to make a model. How do people get the meta data for 40 samples? isn’t that a pretty expensive and difficult proposition? They would need lab test reports on all these 40 samples?
As I mentioned, since choosing the most successful pre-processing method is a trial and error process it is difficult to recommend using specific one. “Processed” is mainly useful for estimation models in which we assume the Beer-Lambert equation holds. “Normalized” is useful mainly for classification models. Sometimes use of both “processed” and “normalized” is needed when there is variance in the optical path of the samples.
We recommend trying each of the methods and evaluate the model’s performance.
Ayelet
Thanks. Can you answer the question below?
Also, i saw that i need at least 40 samples to make a model. How do people get the meta data for 40 samples? isn’t that a pretty expensive and difficult proposition? They would need lab test reports on all these 40 samples?
- You must be logged in to reply to this topic.