Data Driven Calibration of Color-Sensitive Optical Sensor by Supervised Learning for Botanical Application

Color is a fundamental and informative element in botanical applications, as it is in our various daily activities. The conventional ways of evaluating colors based on human visibility are person-specific and thus likely to be biased, whereas color sensing equipment with high sensitivity and accuracy is often a very expensive product. Recently, nonlinear data conversions using machine learning, such as supervised learning, have become available in various fields, and such techniques may be useful for calibrations of low-cost, poor-accuracy sensors. In this study, we propose a method for color-value calibration applicable to low-cost sensors by supervised learning. Specifically, after a supervised learning procedure using training data based on a large number of color samples positioned on a two-dimensional color plane, we improve calibration accuracy over mixed-color areas containing red, green and blue values. We apply this calibrated sensor to estimating the ripeness level of Japanese pears, and the results indicate that the chroma shown in a color coordinate is a good measure for this purpose. In another botanical case, temporal changes in color values of autumn tree leaves in chroma and hue could be successfully detected. This technology is aimed at use with low-cost consumer electronic devices, and it is applicable to color sensing in other areas beyond in the botanical field, thus showing high potential for other kinds of low-cost sensors.