Moisture detection of single corn seed based on hyperspectral imaging and deep learning

Abstract

Precision seeding technology for corn puts forward higher quality requirements for single-kernel corn seed. Moisture is one of four mandatory inspection items of China's national crop seed quality standards. Therefore, accurate determination of the moisture content of a single corn seed is of great significance for corn precision seeding. However, traditional moisture detection methods for corn specified in the national standard, such as the oven-drying method, have the disadvantage of being time-consuming, destroying samples and inability to detect single-kernel samples. The study aimed to explore a rapid, non-destructive, high-precision method to detect the moisture content of single corn seeds based on hyperspectral imaging technology. The hyperspectral images of embryo and non-embryo surfaces of single-kernel samples were acquired with the rage of 968.05–2 575.05 nm. The spectra of each kernel were calculated by averaging the two sides' spectra. The moisture models of single kernels were established separately by PLS, CNN, LSTM and CNN-LSTM based on two sides' spectra. The comprehensive index (R/(1 + RMSE)) was proposed combined correlation coefficient (R) with root mean square error (RMSE) to evaluate the models. The results indicated that the CNN-LSTM model of the embryo-side was optimal for moisture content determination with the comprehensive index (RMSE/(1 + R)) of 0.141 in prediction sets. Therefore, hyperspectral imaging technology combined with deep learning can be considered a promising tool for non-destructive and high-throughput moisture detection in single corn seeds.