Prediction of sugar content in Java Plum using SW-NIR spectroscopy with CNN-LSTM based hybrid deep learning model

Abstract

Sugar content is the most important parameter for consumer acceptance and post-harvest management of Java Plum (Syzygium cumini L.). Traditional methods for sugar content analysis are often time-consuming and labor-intensive. The short wave-near infrared (SW-NIR) spectroscopy offers a rapid and non-destructive alternative for assessing sugar content in fruits. In this study proposed a Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) based hybrid deep neural network model and partial least square regression (PLSR) algorithm for predicting sugar content in Java Plum using SW-NIR spectroscopy. The proposed model combines the strengths of CNN-LSTM to capture sequential dependencies of SW-NIR data in the ranges of 900–1700 nm. The hybrid model is made of 17 layers of neural network and consists of 1D-CNN, LSTM, GroupNormalization and Regularizer layers. At first, the spectra data was preprocessed using several preprocessing techniques independently and developed PLSR model to select the best preprocessing technique. The Savitsky-Golay 2nd derivative preprocessing spectra yielded the most optimum result for PLSR model with coefficient of calibration $R_{cal}$ = 0.677 and coefficient of prediction $R_{pred}$ = 0.554 The proposed CNN-LSTM-based hybrid deep learning model showed the $R_{cal}$ = 0.843 and $R_{pred}$ = 0.83. The results demonstrated the potential of SW-NIR spectroscopy combined with CNN-LSTM-based hybrid deep learning model for determination of soluble sugar content in Java plum.