Hyperspectral fluorescence imaging for mineral classification

In all production industries--including mineral sorting--there is a common aim of providing high-quality results at low cost. Industrial-scale mineral sorting is a particularly sophisticated task. This is because minerals are highly variable natural materials. Conventional mineral sorting techniques--which are based on color images--commonly fail when minerals with similar optical characteristics have to be sorted.

Other mineral sorting methods include laser-induced breakdown spectroscopy (LIBS) and near-IR hyperspectral imaging (NIR-HSI). Such methods can be used to provide high-quality sorting, but they also have several drawbacks. LIBS is a very precise analytical method, but the rate of spectra acquisition is limited.¹ In addition, the laser beam must be pointed to a specific point of interest, which requires substantial technical effort. With NIR-HSI techniques, reflectance images are generally used. At each wavelength a complete grayscale image of the scene is acquired. The images at every wavelength are then concatenated to create a so-called data cube (see Figure 1). A complete spectrum is therefore available for each image pixel. As such, NIR-HSI can be used to provide spatial information about the chemical composition of the imaged scene. The high cost of near-IR sensors (silicon sensors cannot be used because of their large band gaps), however, and the high noise levels of the images, are problems associated with this technique.

Figure 1. Illustration of the structure of a hyperspectral image. For reflectance images, the two image dimensions (x and y), together with the spectral dimension—in this case the emission wavelength (λ_em)—form a so-called data cube. During the acquisition of fluorescence images only one excitation wavelength (λ_exc) is used and the response of the sample to this illumination is measured. The spectral characteristics of the light that is emitted from the sample can depend on λ_exc. When several different excitation wavelengths are used, a data cube is generated for each λ_exc. For hyperspectral fluorescence images with variable illumination, the λ_exc therefore provides a fourth dimension, and multiple data cubes are created.

We have developed a new hyperspectral fluorescence imaging technique that can be used for the industrial sorting of minerals. In our approach, we acquire hyperspectral images in the visible light wavelength range when the spatial resolution of such images is required for mineral identification purposes. This means that we can use relatively cheap and low-noise silicon devices. We are also able to use a UV lamp as a light source rather than a near-IR lamp. In the fluorescence process, UV photons are absorbed in the mineral and cause the emission of visible light photons. The emitted photons carry less energy than the UV photons because of internal loss processes. This change in energy can be used to characterize the mineral species.

Although our new method presents some benefits over traditional HSI techniques, it also shares some of the general disadvantages, e.g., the high data volumes that are produced, and the relatively low spatial resolution compared with conventional color images. An additional weakness of our approach is that specific minerals—depending on their precise chemical composition—can fluoresce at different colors.² This is because fluorescence is caused by luminescent centers, which are often impurity ions in the host mineral lattice. Relatively small proportions of impurities lead to substantial levels of fluorescence photon emission. As there are many different impurities that can cause fluorescence, the measured light is a superposition of the emission from various impurities. In general, fluorescence spectra have a wider shape and exhibit less distinctive spectral characteristics (i.e., peaks and troughs) than near-IR reflectance spectra. Fluorescence spectra are therefore sometimes difficult to assign to the correct mineral species. It is still possible, however, to apply spectral unmixing techniques³ to the acquired spectra. An additional difficulty with our fluorescence technique is that not all minerals fluoresce when they are illuminated with a normal intensity of UV light. We overcome this issue, however, by using light sources with the appropriate power levels (with sufficient power, many minerals will exhibit fluorescence).

Figure 2. Average fluorescence intensity per pixel for a dolomite sample, as a function of λ_em and λ_exc.

We have conducted a set of experiments to assess our hyperspectral fluorescence imaging technique and to determine if it can be used for the successful identification of minerals. In our experiments we used a tunable monochromatic light source (a 300W xenon arc lamp and a 300mm Czerny-Turner monochromator) for illumination. We acquired the hyperspectral images with an acousto-optical tunable filter (a Gooch & Housego HSi-300) in combination with an Andor iXon3 897 single-photon detection electron multiplying CCD camera. We varied the excitation wavelength of the illumination from 280 to 360nm (in steps of 20nm), and we used a channel bandwidth of 15nm. To acquire the hyperspectral fluorescence images, we used a wavelength range of 450–790nm (in steps of 4nm). In this case, the bandwidth of each channel was 3.9nm.

An example of the data we acquired—for a sample of dolomite (an anhydrous calcium magnesium carbonate)—is shown in Figure 2. The average intensity of all pixels in our acquired images is plotted as a function of the excitation and the emission wavelengths. We find that the measured intensity of the light emitted by the sample is dependent on the excitation wavelength that is applied and the emission wavelength. A full description of our experimental results is available elsewhere.^4,5

We have developed and experimentally verified a new hyperspectral fluorescence imaging technique that can be used for mineral identification and industrial sorting purposes. The results we have obtained so far are encouraging, and we continue to refine our approach. Our ultimate goal is for fluorescence acquisition and classification methods to be implemented within practical mineral sorters. An example of an industrial-scale sorting problem is the upgrading and pre-concentration of magnesite (a fluorescent mineral). In our current work we are focusing on making adequate data preparations. For instance, we wish to properly remove noise from our images so that we can achieve high mineral classification rates. In addition, we are working on performing mineral classification by considering the hyperspectral images that are acquired at a number of different excitation wavelengths (rather than just one). We are also considering other future developments, e.g., using more narrow wavelength regions that are sufficient to make correct mineral identifications. This would reduce the amount of data that is acquired and processed in our technique.