Let’s get into the data that we collected with our hyperspectral cameras. When we got samples from the cement and steel production, we measured them in our lab, placing each sample side by side and scanning them with our cameras.

Just to remind you, we are scanning in wavelength ranges that include wavelengths in the visible light, the same color information that we can see with the naked eye, but our hyperspectral cameras far surpass this and are able to detect longer wavelengths up to 2500nm. In these wavelength ranges we can see material changes that tells us information about the chemistry (for example the mineral content or iron content) or the physical characteristics of the surface (for example the grain size).

Though this is not a direct measurement, the reflected light from the surface can be used to indicate this information. Before looking at the actual original and used brick products that are being used in cement and steel production, we wanted to look at the input raw materials that go into the brick production to get an idea about their spectral “fingerprint”. Doing that is important, even though the bricks themselves change a bit in composition because of the heat and pressure that they experience during use and the different gases and fumes that affect them and cause e.g., changes in mineral composition.

For the steel casting ladle bricks, the main input raw materials are plotted in the visible to near-infrared wavelengths region below (400 – 1000nm, abbreviated as “VNIR”).
You can see that the plotted lines (spectra) show some differences. For example, the two different MgO components, fused (green) and sintered (blue) look very different from each other.

In the second step, we wanted to try to classify the different materials by training a model on the existing data and predicting the class of each pixel in the image. The table below shows the prediction of each pixel (pixel-based). It also shows the class prediction of the raw material for the whole image that is either based on counting the pixels and assigning a class based on the dominant class of all pixels (pixel-class majority) or creating an average spectrum from all pixels in the image and predicting the class based on this one spectrum (average spectrum).

You can see that for this smaller subset of materials, the prediction of raw materials seems to work rather well using VNIR-data, even though not all pixels are classified perfectly (e.g. see the diverse pixel classification in Bauxite 3-6mm)