Research On Fine Classification Of Rice Varieties In Cold Region Based On Double-scale Hyperspectral Information

Rice is an indispensable food crop for human beings.More than 60% of the population in China relies on rice as their staple food.Rice varieties are closely related to disease resistance,insect resistance,lodging resistance,grain quality and yield.Government policy makers need to select suitable rice varieties based on land and environmental conditions to increase crop yields and land use,thereby creating higher economic returns and promoting sustainable agricultural development.Growers need to implement precise fertilization and irrigation techniques for different rice varieties to reduce the pressure and pollution caused by excessive fertilization on the ecological environment.National agencies need to know the distribution and characteristics of rice across the country,and count information such as the varieties and yields of rice planted.However,different varieties of rice have similar appearance traits and the same growth trend,relying on the knowledge and experience of agricultural experts to analyze and identify,there is subjectivity,high error rate,and cost manpower,material resources and time.Traditional identification methods can not be used to analyze rice quickly,accurately and in real time.Therefore,from the perspective of intelligent agriculture and precision agriculture,this study adopts hyperspectral technology with characteristics of fast and non-destructive,takes hybrid rice in cold region as the research object,and carries out analysis and research on accurate classification of rice varieties based on hyperspectral data of ground and UAV scales.The research contents and results mainly include:(1)On the scale of UAV,the UAV equipped with S185 hyperspectral equipment was used to acquire the hyperspectral image of rice varieties,and the data set and label of rice varieties were made.After principal component analysis was used to reduce the dimensionality of the data,a hybrid convolutional neural network was used to automatically learn the spectral and spatial characteristics of 14 rice varieties and deep extract them.In addition,to further improve the performance of the classification model,we tried to optimize it with an end-to-end trainable attention module.Finally,extensive experiments were used to prove the validity of the model.Compared with six advanced machine learning methods,the 3D-CSAM-2D-CNN model proposed in this paper achieved the best classification effect on fine classification of rice varieties.The overall classification accuracy of the model was 98.93%,and the classification accuracy of single rice variety was over 98.22%.(2)On the ground scale,portable equipment was used to collect the canopy hyperspectral data of rice at the boosting stage,and data set containing only spectral reflectance information and label were made.In order to realize fine classification of rice varieties,a one-dimensional convolutional neural network model based on self-attention mechanism was proposed.Firstly,five preprocessing methods(Savizky-Golay smoothing filter,standard normal variate transformation,multiple scattering correction and two combined preprocessing methods)and three extraction and selection methods(principal components analysis,competitive adapative reweighted sampling and successive projections algorithm)were used for data processing.Then,three classification models(K-nearest neighbor model,random forest model and the model proposed in this study)were used to compare the classification effect,and the "preprocessing-feature extraction-classification model" with the best effect was selected to complete the high-precision classification task of rice varieties.Experiments showed that the model proposed in this paper(Self-Attention-1D-CNN)has the highest classification accuracy,with an average classification accuracy of 99.93% in ten experiments.Based on double-scale hyperspectral data,this study proved the feasibility of hyperspectral technology for fine classification of rice varieties and the convolutional neural network model is a potential classification method for high-precision classification of crops.At the same time,the model proposed in this research is beneficial to the automatic identification of field crops and the monitoring of near-ground agricultural conditions,which contributes new possibilities for crop phenotype research and the implementation of precision agriculture.