Related Experiment Video
Updated: Jul 17, 2026
Automated Modular High Throughput Exopolysaccharide Screening Platform Coupled with Highly Sensitive Carbohydrate Fingerprint Analysis
Published on: April 11, 2016
A multi-task deep attention network for simultaneous rapid quantification of sucrose, glucose, and fructose contents
Yingchao Xu1, Jiayu Luo1, Shudan Xue1
1Guangdong Key Laboratory for New Technology Research of Vegetables, Vegetable Research Institute, Guangdong Academy of Agricultural Sciences, Guangzhou 510640, China.
Abstract:
Sugars critically influence pumpkin quality, however, quantifying them using Fourier transform near-infrared (FT-NIR) spectroscopy remains challenging due to spectral overlap. This study developed a Multi-Task Deep Attention Network (MTDAN) integrating deep architecture, attention mechanisms, and multi-task learning for simultaneous quantification of three major soluble sugars. Pumpkin FT-NIR analysis showed MTDAN outperformed partial least squares regression, Ridge regression, and random forest. MTDNA achieved superior prediction accuracy (coefficient of determination of prediction R2p = 0.91-0.93 and a root mean square error of prediction RMSEP = 8.38-10.30 mg/g DW), and wide concentration ranges of fructose (20.18-139.68 mg/g DW), glucose (16.52-157.73 mg/g DW), and sucrose (13.54-209.64 mg/g DW), along with robustness to spectral overlap. Band-specific experiments revealed that localized spectral regions (4000-4800 nm for fructose and glucose, 5600-6400 nm for sucrose) contained actionable signals. This study established MTDAN as a versatile tool for accurate quality assessment of sugars in pumpkins.
