Estimation of soluble solids content and fruit temperature in 'rocha' pear using Vis-NIR spectroscopy and the spectraNet–32 deep learning architecture

• Main Author: Martins, J. A
• Publication Date: 2023
• Other Authors: Rodrigues, Daniela, Cavaco, A. M., Antunes, Maria Dulce, Guerra, Rui Manuel Farinha das Neves
• Format: Article
• Language: eng
• Source: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Download full: http://hdl.handle.net/10400.1/19722
• topic: Deep learning; Residual network; Near-infrared; Spectroscopy; Pear fruit

Spectra-based methods are becoming increasingly important in Precision Agriculture as they offer non-destructive, quick tools for measuring the quality of produce. This study introduces a novel approach for esti-mating the soluble solids content (SSC) of 'Rocha' pears using the SpectraNet-32 deep learning architecture, which operates on 1D fruit spectra in the visible to near-infrared region (Vis-NIRS). This method was also able to estimate fruit temperatures, which improved the SSC prediction performance. The dataset consisted of 3300 spectra from 1650 'Rocha' pears collected from local markets over several weeks during the 2010 and 2011 seasons, which had varying edaphoclimatic conditions. Two types of partial least squares (PLS) feature selection methods, under various configurations, were applied to the input spectra to identify the most significant wavelengths for training SpectraNet-32. The model's robustness was also compared to a similar state-of-the-art deep learning architecture, DeepSpectra, as well as four other classical machine learning algorithms: PLS, multiple linear regression (MLR), support vector machine (SVM), and multi-layer perceptron (MLP). In total, 23 different experimental method configurations were assessed, with 150 neural networks each. SpectraNet-32 consistently outperformed other methods in several metrics. On average, it was 6.1% better than PLS in terms of the root mean square error of prediction (RMSEP, 1.08 vs. 1.15%), 7.7% better in prediction gain (PG, 1.67 vs. 1.55), 3.6% better in the coefficient of determination (R2, 0.58 vs. 0.56) and 5.8% better in the coefficient of variation (CV%, 8.35 vs. 8.86).