Identification Of Rice Growth Stage Based On Fractal Dimension And Deep Learning

Intelligent agriculture is a new model of agricultural development appearing in modern information technology,which is a new direction for our country’s agriculture to automation and modernization.In the automatic cultivation of rice,different growth stages correspond to different cultivation methods.The traditional identification methods of rice growth stages mainly rely on manual observation.However,manual observation methods are tedious,subjective and prone to errors,especially in the context of large-scale rice cultivation.The utmost importance of using a computer to accurately determine the stage of rice is thus clear.On the one hand,this paper calculates a rice texture parameter,fractal dimension,according to the rice image,and proves that this parameter can effectively help model recognition.On the other hand,this study uses a method that combines machine learning with deep learning,especially with convolutional neural networks,and the result is demonstrated to be more advantageous than traditional machine learning techniques.The extraction of a fractal dimension and the combination of a mixed model enabled the examination of the automatic recognition of rice growth stages.Firstly,the rice image and data obtained by the smart agriculture team of Huazhong Agricultural University were used to analyze various rice characterization parameters,and the fractal dimension was calculated using the rice image.The data set and image set of rice at three growth stages were obtained,including tillering stage,jointing stage and heading stage.Then the threshold method was used to convert these images into binary graphs,and the box counting method was used to calculate the fractal dimensions of IFD and SFD.The heterogeneity analysis and normality test proved that the fractal dimensions could be effectively used for classification.Statistical methods,random forest feature screening method and Garson method based on neuron sensitivity were comprehensively considered to form rice parameter data set.Secondly,machine learning and deep learning were used to identify the growth stages of rice.On the one hand,five kinds of machine learning classifiers were used to identify the growth stage of rice with the selected data set and fractal dimension as input features,and a ten-fold cross-validation technique was utilized to assess the model’s impact.On the other hand,a convolutional neural network was constructed to identify the development stage of rice color pictures,and the self-coding network was contrasted with vgg16-net.And,the class activation mapping method was employed to visualize the output of convolutional layer interneurons,and the decision basis for the convolutional neural network to extract rice features was then examined.Finally,the Convolutional Neural Networks-Machine Learning(CNN-ML),a hybrid model of deep learning and Machine Learning,is applied the fractal dimension to identify the growth stage of rice.Aiming at the disadvantages of complex feature extraction and single deep learning classifier,this paper combined rice features automatically extracted from convolutional neural network with fractal dimension as input features of machine learning classifier for growth stage recognition.Through experiments,for support vector machine,the accuracy of the model using only rice data set is 82.63%,the accuracy of the model with IFD fractal dimension is 88.02%,and the accuracy of the CNN-ML model with fractal dimension is94.52%.