Measurements of Anthocyanin Content of Prunus Leaves Using Proximal Sensing Spectroscopy and Statistical Machine Learning

A novel measurement framework for absolute total anthocyanin content (AAC) in leaves is proposed, where proximally-sensed leaf spectroscopic measurements are coupled with machine learning techniques. Use of both leaf reflectance and transflection spectral measurements is investigated by sounding the entire spectral range, from the visible to the short-wave infrared, and accounting for the whole shape of the leaf spectra. Results show low prediction errors (not higher, on average, than 0.34 mg/g) for AAC retrieval even when only a quarter of available data are used for training, with improved performance for higher fractions of training data. The employment of transflectance is shown to benefit to the retrieval process, thus providing generally lower prediction errors. During the AAC retrieval process, the machine learning technique automatically performs wavelength selection. By providing as output the subset of most relevant wavelengths for AAC retrieval, this framework represents a first step toward the development of a low-cost measurement system to be easily operated on the field for quick, reliable anthocyanin content measurements.