Accurate prediction of kiwifruit soluble solids content using an optimized 1D-CNN model with near-infrared spectroscopy

Abstract

With the growing demand for ready-to-eat kiwifruit in the world, consumers are placing higher demands on the edible quality of kiwifruit, particularly its soluble solids content (SSC), which determines its flavor and internal quality. Therefore, rapid, accurate, and nondestructive assessment of kiwifruit SSC in the sorting and supply chain stages is crucial for enhancing product value and market competitiveness. Traditional methods for measuring sugar content are often time-consuming and may damage the sample, while Deep Learning (DL) technology has shown significant potential in many fields nowadays. However, its application to the nondestructive prediction of kiwifruit sugar content remains limited. This study focuses on the “Cuixiang” kiwifruit, collecting its near-infrared (NIR) spectral data within the wavelength range of 996–1710 nm. Subsequently, we propose a one-dimensional convolutional neural network (1D-CNN) model and compare its performance with four traditional machine learning models, i.e. SVR-RBF, SVR-Linear, SVR-Poly, and PLSR. Furthermore, we evaluate the optimization effects of various spectral preprocessing and feature selection methods (such as SNV-CARS5) on model accuracy. The results show that the 1D-CNN model demonstrates exceptional predictive capability. On the raw spectral data, its (Formula presented.) and (Formula presented.) are 0.8034 and 2.3973, respectively, both of which are significantly better than traditional models. After data processing and feature selection strategy, the model performance reaches its peak, with (Formula presented.) and (Formula presented.) as high as 0.9387 and 3.9256, respectively. Overall, this paper provides a reliable theoretical foundation and technical support for the nondestructive, accurate, and intelligent prediction of kiwifruit sugar content.