Identification Of The Rice Growth Stages Based On The BP-Adaboost And The Fractal Dimension

At present,agricultural development is mainly oriented towards precision agriculture and intelligent agriculture,and the effective identification of crop growth stage is an important application of these two directions in agricultural development.However,there is still a lack of research on the application of crop image features to crop growth stage recognition.In this study,a rice texture parameter named fractal dimension is calculated according to the change of green leaf area and rice image,and the experiment prove that the parameter could effectively improve the classification effect of the model.Based on the extracted fractal dimension,a method and system for rice growth stage recognition based on machine learning and fractal dimension is proposed in this paper.Firstly,an automatic rice phenotypic RAP system developed by the smart Agriculture team of Huazhong Agricultural University is used to obtain the phenotypic characteristic data set of rice,and the image data are preprocessed.RAP system is used to obtain phenotypic characteristic data sets and color images of rice at three growth stages,including tillering stage,jointing stage and heading stage,and the phenotypic characteristic data sets are screened.Furthermore,the collected rice color images are grayed,and the gray scale of rice is obtained by B channel graying.The rice image is binarized by Ostu segmentation method and denoised by expansion and corrosion operation.In addition,considering the scale-free and autocorrelation of rice images with increasing growth,fractal dimension characteristics of rice images are extracted and numerically analyzed.For the rice binary image,the box-counting dimension method is used to calculate the two fractal dimensions based on the whole rice and the enclosing rectangle under the binary graph,which are denoted as DBCS and SFD respectively.Then for the gray scale image of rice,box-counting dimension method and random walk method are used to calculate the two fractal dimensions based on the whole rice in the gray scale,which are denoted as DBCG and FBC respectively.According to the normality of the fractal dimension with the growth of rice and the botanical theory,the significance of adding the fractal dimension in this study is given.Finally,feature screening and classifier construction are carried out for rice phenotypic data sets,and the structure of the original Adaboost integrated model is optimized.All the features in the original rice data set are tested for significance at two stages,and the insignificant features in two stages are removed.Then,random forest feature selection method in feature selection tree model is used for feature selection.The contribution degree of each feature to rice growth stage category is calculated,and the top eight features are retained.Six traditional classifiers are used to construct the classification model,and BP neural network is used to replace the weak classifier CRAT decision tree of the traditional Adaboost ensemble classifier.Finally,after cross validation and bayesian optimization,the accuracy rate reach97.33%,the accuracy reach 94.39%,and the F1 value reach 0.94.