Assessing Coffee Bean Roasting Time with Regression Analysis of Hyperspectral Data

Whether precisely monitoring coffee bean roasting time for the perfect cup of coffee or detecting an emerging algae bloom on a lake, the spectral sensitivity of Resonon’s hyperspectral cameras can quickly provide answers to manufacturers and researchers alike.

As remarkable as human vision is, it is based on only 4 types of sensors on our retina (rods and three types of cones). We effectively “see” only three colors and our range of sensitivity is limited to light of wavelengths between 400 and 700 nm. Our eyes can qualitatively observe coffee beans transforming from greenish to reddish and from light to dark during the roasting process (see Figure 1), but we are unlikely to make accurate quantitative assessments by eye.

Precise record-keeping and process control ensure that coffee roasters repeatedly achieve the perfect roast and blends for the best cup of coffee. A mistake or fault in the process can lead to under or over-roasted beans, which translates into wasted time, product, and money (and possibly a lousy cup of Joe). This study illustrates how the roasting time of coffee beans can be determined in real-time with hyperspectral imaging.

We scanned 8 samples of coffee beans roasted for 0, 2, 4, 6, 8, 10, 12, and 14 minutes, respectively. We used seven of those scans (all except the 12-minute roasted beans) to train a regression model for roasting time, reserving the 12-minute roasted beans data to test the models accuracy.

For those unfamiliar with regression analysis, it is a statistical technique that quantifies the relationship between one or more independent variables and a dependent variable. In this case, the regression model quantifies the relationship between the average reflectance spectra of a datacube of roasted coffee beans (the independent variable) with the time those coffee beans were roasted (the dependent variable). After training, the regression model can predict how long a batch of coffee beans was roasted based on the average spectral signature of those beans. Figure 2 shows the average spectral signature for each group of beans shown in Figure 1.

After training the regression model on all the samples except the 12-minute roasted sample, the regression model was asked to predict how long the 12-minute beans had been roasted. The model estimated 12.1 minutes.

Even though the bean color varies greatly within a batch, and even across a single bean, the algorithm can crunch the data from millions of pixels (which, in this example, each contain 447 color channels) to extract meaningful information and estimate the roasting time for every single pixel in the scan. Then, averaging over all pixels in the scan reveals an accurate prediction of the batch’s roasting time.

This simple study illustrates how non-destructive measurements with Resonon’s hyperspectral imaging system allow you to gain valuable insight into processes that are otherwise challenging to monitor.

The sample data used in this study are available for free at our Downloads site. 

Thank you to our friends at Ghost Town Coffee Roasters for providing the roasted beans for this study!