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

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