INTELLIGENT IDENTIFICATION OF RICE GROWTH PERIOD (GP)BASED ON RAMAN SPECTROSCOPY AND IMPROVED CNN IN HEILONGJIANG PROVINCE OF CHINA

ABSTRACT

The fertile land in Heilongjiang Province of China is suitable for rice cultivation, but this area is susceptible to low temperature and chilling injury, which is prevented by planting rice varieties suitable for GP that is an important measure. However, selection based on rice traits is vulnerable to environmental influences and takes a long time, and selection based on molecular markers may result in progeny recombination and lack of reliability. Therefore, an efficient accurate and intelligent identification method for rice growth period is urgently needed. In this study, machine learning and deep learning methods in Python were used to analyze the Raman spectra of 6 rice varieties in three accumulated temperature region of Heilongjiang Province. 1) In machine learning, Principal Component Analysis (PCA) was adopted for feature extraction, in combination with Support Vector Machine (SVM) classification models suitable for nonlinear data sets for identification, the identification rate was 93.33% and the type of this experimental data set was determined to be discrete. 2) In deep learning, Continuous Wavelet Transform (CWT) methods was adopted for data preprocessing, combined with the Convolutional Neural Networks (CNN) model with its own feature extraction, with the highest accuracy of 94.82%, which was higher than the PCA+SVM identification model. 3) Based on the method mentioned in 2), in order to improve the feature extraction ability of the model as a whole, Convolutional Block Attention Module (CBAM) was used to improve the CNN identification model for the first time for one-dimensional data sets, and the highest identification rate was 98.28%, which was better than the PCA+SVM identification model. 3) In the verification test, Raman spectral information of 4 rice varieties was brought into the constructed CWT+CNN-CBAM identification model for identification, and the identification results were as high as 94.79%. The experimental results showed that the CWT visualization data processing method based on Raman technology combined with the CNN identification model of CBAM with improved feature extraction ability in deep learning achieved the best identification results, which could provide an efficient, accurate and intelligent method for the identification of different growth period of rice varieties in Heilongjiang Province.

rice growth period; Raman spectroscopy; low temperature; chilling injury; CNN-CBAM