Classification of the maturity stage of coffee cherries using comparative feature and machine learning

Abstract

This work presents the use of multiple techniques (i.e., physicochemical and spectral) applied to harvested coffee cherries for the postharvest classification of the maturity stage. The moisture content (MC), total soluble solids (TSS), bulk density, fruits’ hardness, CIEL*a*b parameters and the dielectric spectroscopy methods were applied on coffee cherries at seven maturity stages. These maturity stages were assessed according to the days after flowering (DAF) and the physical appearance as traditionally performed by growers. An increase of the green-to-red ratio (i.e., a*) parameter was perceived, accompanied by a monotonic response for the hardness, TSS and bulk density with a maximum moisture content at stage 5. In the case of the dielectric spectroscopy technique, the loss parameter presented higher losses for unripe stages at the ionic conduction region. To compare the individual performance of each of the techniques, three machine learning methods were used: random forest (RF), support vector machine (SVM) and k-nearest neighbours (k-NN). The meta-parameters for these techniques were optimized for each case to achieve the best performance possible. Furthermore, as the dielectric response is of spectral nature, recursive feature selection was applied and the 500 MHz to 1.3 GHz frequency range selected for the task. The highest performance was obtained for the colorimetric (75.1%) and hardness (72.5%) responses, while the lowest was obtained for the moisture content (45.5%). The dielectric spectroscopy response presented a promising response (56.8%), that achieved a clear separation of unripe from ripe stages, except for stage 5 in which some of the samples were classified as stage 2. Most techniques studied are compatible with field conditions, and the dielectric technique shows potential to be transferred based on available software-radio defined platforms.