Multi-sensor spectral fusion to model grape composition using deep learning

Abstract

Spectral instruments can be useful for the rapid assessment of chemical compounds in different targets, and their use have been already reported for the modeling of grape composition comparing two spectral ranges. Still, with the increased easiness of acquiring data with several sensors, it would be valuable to explore spectral fusion techniques for the modeling with deep learning, seeking to obtain improved performance. Therefore, the objective of this work was to develop multi-sensor spectral fusion approaches for the deep learning modeling of grape composition. From 128 grape samples, two spectra per sample were acquired from two different ranges using two sensors (visible and shortwave near infrared, 570–1000 nm; and wider NIR 1100–2100 nm). From each sample, 15 grape nitrogen compounds were analyzed by wet chemistry. Three different data fusion approaches are defined using neural networks and deep learning, testing several ways of structuring and merging the input spectra. Statistical analyses supported that (i) the proposed deep learning fusion architectures performed better than single spectral range models, and (ii) neural networks have better modeling capabilities than partial least squares in spectral fusion. The results demonstrate the potential of deep learning for spectral data fusion in grape nitrogen composition regression, and potentially other traits in food and agriculture spectroscopy.