SpectraNet-53: A deep residual learning architecture for predicting soluble solids content with VIS-NIR spectroscopy

Abstract

This work presents a new deep learning architecture, SpectraNet-53, for quantitative analysis of fruit spectra, optimized for predicting Soluble Solids Content (SSC, in °Brix). The novelty of this approach resides in being an architecture trainable on a very small dataset, while keeping a performance level on-par or above Partial Least Squares (PLS), a time-proven machine learning method in the field of spectroscopy. SpectraNet-53 performance is assessed by determining the SSC of 616 Citrus sinensi L. Osbeck 'Newhall' oranges, from two Algarve (Portugal) orchards, spanning two consecutive years, and under different edaphoclimatic conditions. This dataset consists of short-wave near-infrared spectroscopic (SW-NIRS) data, and was acquired with a portable spectrometer, in the visible to near infrared region, on-tree and without temperature equalization. SpectraNet-53 results are compared to a similar state-of-the-art architecture, DeepSpectra, as well as PLS, and thoroughly assessed on 15 internal validation sets (where the training and test data were sampled from the same orchard or year) and on 28 external validation sets (training/test data sampled from different orchards/years). SpectraNet-53 was able to achieve better performance than DeepSpectra and PLS in several metrics, and is especially robust to training overfit. For external validation results, on average, SpectraNet-53 was 3.1% better than PLS on RMSEP (1.16 vs. 1.20 °Brix), 11.6% better in SDR (1.22 vs. 1.10), and 28.0% better in R² (0.40 vs. 0.31).

Plant phenotyping relevance

果実スペクトルから糖度という植物器官形質を推定する深層学習アーキテクチャを開発し、既存手法と内部・外部検証で比較しており、表現型取得・推定法が中心である。

Code and data availability

公開状態または取得可能な本文経路を確認できませんでした。

No evidence-backed public reproduction asset is currently recorded.