Feasibility study on identifying seed viability of Sophora japonica with optimized deep neural network and hyperspectral imaging

Abstract

Aiming at the rapid identification of seed viability of Sophora japonica, different aging grades and imbibition states were proposed as the research focus. The hyperspectral data of the four viability grade seeds in the state of swelling at 0 h and 10 h were collected. The effects of swelling and aging on the content of water, protein, starch, fat and total sugar were compared, and then correlation analysis was established. Among them, fat was the most sensitive to changes in spectrum and viability. In the machine learning scale, naive bayes (NB), self-organizing feature mapping algorithm (SOFM), and support vector machine (SVM) were used. On the scale of deep learning, convolutional neural networks (CNN) and long short-term memory (LSTM) were applied to model on the one-dimensional spectral and two-dimensional image. The image-based deep learning model performed best among the three models, and the recognition accuracy was above 90%. A particle swarm optimization algorithm (PSO) was proposed to optimize the hyperparameters. It can increase the recognition rate, up to 99.73%, and increase the convergence speed. This study proved that deep learning combined with spectral or image has great potential in identifying the seed viability.