Very often, when discussing with customers, we need to explain the differences between hyperspectral cameras from multispectral ones. A standard definition states that hyperspectral cameras have more than 100 bands, whereas multispectral ones have fewer. But this definition does not consider the width of the spectral range or the sampling. That means that if a camera covers the spectral range 400 – 600 nm with 50 bands, it would not be hyperspectral, whereas if it covered 400 – 800 nm with the same sampling (meaning this time 100 bands), it would be hyperspectral. We find this discrepancy not relevant and prefer to speak about spectral resolution (FWHM, Full Width Half Maximum*), highlighting the ability of a camera to separate two consecutive spectral peaks from each other.

*The full width half maximum defines the spectral resolution of a system, more precisely, in this context, its ability to separate spectral peaks from each other.

In this (not perfect) attempt at definition, we consider that a hyperspectral camera provides smooth and most resolute spectra. In contrast, the ones supplied by multispectral devices are more like stairs or saw teeth-like figures without abilities to depict acute spectral signatures.

We will study the sorting of shells among almonds to illustrate this. This is a very typical application, and this case study echoes another one we already published on our website (Hyperspectral Technology vs. RGB). One can say that an RGB camera is a multispectral device to some extent.

We first made a review on multispectral cameras available on the market. The spectral range is limited to 400 – 1000 nm for most of them, and the number of bands is often 4 or 5. Those are crucial limitations for many applications.

1. Limit on the spectral range

Reflection, absorption, or emission features shapes spectra tightly related to the molecular composition of scrutinized material. The below table is well known and highlights the spectral bands where each of the most common molecules has their electromagnetics resonance overtones. As can be seen, the spectral range of 700 – 2500 nm is necessary for many applications. Especially for those related to food quality assessment and plastics sorting, the range of 1100 – 1700 nm is mandatory. Consequently, it highlights that cameras limited to 400 – 1000 nm are not the most relevant for those applications.

Our first study (Hyperspectral imaging vs. RGB cameras) showed that RGB cameras did not perform well in sorting nuts and pistachios. In contrast, the FX10 provided better results, and the FX17 had the best sorting capabilities.

Table 1: Molecular electromagnetics resonance overtones

2. Limit due to the number of bands

Considering the previous point, we made a comparative study on almonds and shells with FX17 data: without and with binning (i.e., merging consecutive spectral bands). The spectral range of 900 – 1700 nm was covered with 224 bands on the first data set, whereas on the second data set, 28 binned bands were used. As can be seen in Figs. 1 and 2, spectra related to the 224 bands dataset are much smoother than those depicted with only 28 bands, and small but crucial spectral differences can be better picked.

Figure 1: spectra of shell and almond depicted by 224 and 28 bands.
Figure 2: FX17 spectrum of almond with 224 and 28 bands. Purple circles highlight the spectral differences due to the presence of oil in almonds but not in the shell.

Besides, with multispectral data, some pre-processing methods are not suitable. For instance, derivative or smoother such as Savitzky-Golay require continuous spectra to perform well, which are not provided by multispectral sensors. We built two models to illustrate these points, as mentioned above, related to the ability to depict acute spectral features. The hyperspectral model (224 bands) is more accurate than the multispectral one (28 bands). The edge effect disappears, and small pieces of shell are not misclassified.

Figure 3: Models based on 28 and 224 bands and these predictions (green for almond, blue for shell).

We selected 28 spectral bands to simulate a multispectral camera. However, typical multispectral cameras can have notably fewer spectral bands, making their ability to depict fine spectral features even lower.

Finally, we acknowledge that custom-made multispectral cameras with well-chosen bands would perform equally well as hyperspectral devices for some applications where the entire spectral range is not needed. However, the user will lose the flexibility offered by hyperspectral sensors (having the possibility to upgrade the machine to sort new types of contaminants or materials). Advantageously, the user can freely select relevant bands with the FX cameras. Therefore, the FX cameras can become multispectral, whereas a multispectral camera could never become hyperspectral. Besides, multispectral cameras are not cheaper, especially if they need to be custom-made, with a relatively large number of bands.