Optimizing Starch Content Prediction in Kudzu: Integrating Hyperspectral Imaging and Deep Learning with WGAN-GP

Abstract

Rapid and non-destructive prediction of starch content in kudzu is essential for the food industry. In this work, we present an approach combining hyperspectral imaging (HSI) and deep learning (DL) techniques for predicting kudzu root starch content. Practical constraints such as equipment and experimental conditions limit the quantity of spectral data and labels obtained, which leads to diminish model performance. To address this restriction, we employ Wasserstein generative adversarial networks with gradient penalty (WGAN-GP) to augment spectral and starch content data simultaneously. Through numerous iterations, synthetic data that closely resemble real data is generated, which is validated through comprehensive evaluations using various qualitative and quantitative analysis. Additionally, we establish and compare partial least squares regression (PLSR), support vector regression (SVR) and one-dimensional convolutional neural network (1DCNN) model before and after data augmentation. Experimental results demonstrate that the introduction of synthetic data could improve model performance significantly. Particularly, 1DCNN model exhibits the best performance, achieving correlation coefficients (R2) of 92.97% and 93.43% for starch content in the two types of kudzu roots. Overall, this study not only provides an effective method for rapidly, non-destructively, and accurately determining starch content in kudzu roots, but also addresses the challenge of requiring a large amount of data.