Abstract: In this paper, hyperspectral technology was used to detect the moisture content in rice. Hyperspectral images of 120 rice samples were collected and preprocessed by multiple scatter correction (MSC). Due to the large amount of original spectral data and their strong redundancy, a stepwise regression (SWR) analysis method was adopted for feature extraction after preprocessing. Finally, a quantitative prediction model for rice moisture content was built based on BP neural network with and without optimization of weight and threshold optimized using genetic algorithm (GA) and mind evolutionary algorithm (MEA), respectively. A comparison was made among BP, GA-BP and MEA-BP prediction models, of which the determination coefficients for the prediction set were all above 0.86. The results showed that the prediction performance of MEA-BP model was the best among these three models with a determination coefficient for the validation set of 0.966 3, and a root mean square error of 0.81%.

Key words: hyperspectral, rice, moisture content, BP neural network, genetic algorithm, mind evolutionary algorithm