Band reduction of the spectral signature for the determination of models based on Machine Learning and spectroscopy using hyperspectral imaging in cocoa beans.

The agribusiness in Peru has grown significantly in the last decade and has become one of the most important producers of fine and flavor cocoa. The European Union has established quality parameters for the presence of heavy metals, one of which is cadmium due to its harmful effects on human health. To improve the quality control of cocoa, this article proposes to estimate the percentage of cadmium in a cocoa sample using the hyperspectral imaging and Machine Learning algorithms, as a non-invasive, non-destructive method that can be implemented in real time. In addition, areas of the spectral signature that contribute the most to the cadmium estimate are identified in order to significantly reduce the number of spectral bands used in the model. This reduction of spectral bands allowed to increase the R2 with the validation data from 67.92% to 72.39%, reaching an average error of 0.18 ppm using the Partial Least Square (PLS) method.