Research On Monitoring Rice Growth Stage Based On Optimized Weighted Intergrated Model

Rice acts a significant role in agricultural production.Precise and timely monitoring of rice growth stage can enable growers to understand variety heterogeneity and adopt appropriate cultivation and management methods according to the physiological properties of rice in various growing periods and the needs of external environment.Currently,the identification of rice growth stages is mostly carried out using manual inspection methods,but there is less research on automatic identification and monitoring of rice growth stages,and traditional research on crop growth stage monitoring relies on morphological related feature data.In this study,ten fractal dimension and gray level co-occurrence matrix texture features are extracted based on rice image calculation,and a new method of rice growth stage recognition and monitoring based on feature modeling using domain knowledge of machine learning is proposed,which can achieve high-precision recognition of rice growth stage with limited samples.The following work has been completed in this paper:First,image pre-processing and preliminary analysis of image characteristics are carried out according to the rice color image provided by RAP system.Firstly,the image is segmented,and the rice plant is separated from the background to obtain the rough image of the rice plant;Secondly,the image is grayed,and for enhancing the image we use the histogram equalization method.The masking algorithm is used to preserve only the rice plant,and for reducing the impact of noise we use the Gaussian filtering method on the image quality;Then the rice plant color in the gray scale image is changed to white to get the rice binary graph;At last,we extract the co-occurrence matrix and the fractal dimension of the pre-processed image,and the newly added features are tested to ensure that they can be used as modeling data sets.Secondly,the rice features obtained from RAP system and ten texture features extracted from rice images are statistically analyzed.Firstly,the Spearman correlation between the input features and category markers and the Spearman correlation between multiple features are analyzed,and the features with correlation lower than a certain threshold are eliminated;Then,we continue to use the random forest recursive feature elimination method for the remaining 35 features.In the iterative process,we eliminate the secondary features according to the weight coefficient of the features given by the random forest model;Finally,30 feature subsets with the highest average score were retained as the rice modeling data set.Thirdly,different single machine learning models and integrated machine learning models are used to identify and classify the rice growth stages,focusing on the analysis of the model effect and the importance of features before and after feature screening,adding fractal dimension and gray level co-occurrence matrix,and using evaluation indicators for comparison.Through comparison,it can be seen that compared with the basic model,the integrated model provides better performance,in which the accuracy and F1-score of optimized weighted integration reach 94.06% and 0.94,and the kappa coefficient is 0.92,it is about 1.57% higher than the basic model without selecting features;The evaluation index of each model has been improved after the introduction of texture features,especially after the addition of gray level co-occurrence matrix,the accuracy of the model has been improved by 2%-7%;The top ten input variables of feature importance include two gray level co-occurrence matrix feature variables extracted from the image,and two fractal dimension variables.In short,this paper systematically explored the application of image processing and machine learning technology in identifying and monitoring rice growth stage.The integrated models have shown great potential in improving the accuracy of rice growth period monitoring,and the fractal dimension and gray level co-occurrence matrix features have important value in rice growth monitoring and growth cycle discrimination.