Research On The Classification Method Of Typical Crops Based On Image Multi-features

Accurate access to the distribution of crop cultivation is important for the country to grasp the macroscopic situation of food production,formulate relevant agricultural policies and forecast the comprehensive production capacity of agricultural resources.Remote sensing technology is widely used in agriculture,forestry,transportation,water conservancy and other industries because it can efficiently acquire large-area features,especially in the field of agriculture,which is an important source of agricultural data acquisition and has an irreplaceable role.Compared with other satellites,the WFV sensor on board the Gaofen-1 satellite has a wider coverage and shorter revisit cycle,which is widely used in agricultural resource survey and dynamic monitoring,crop yield estimation,crop growth and monitoring of pests and diseases.Satellite remote sensing images not only contain self-referential spectral band information,but also have rich spatial relationship characteristics,in crop classification and planting structure extraction,researchers mostly use spectral features or index features for classification and extraction,some scholars also use texture information of images for classification and obtain better classification results,but there are relatively few studies on combining texture features and index features as classification features,and there is relatively little discussion on However,there are relatively few studies on combining texture features with index features as classification features,and there is a lack of discussion on the benchmark window and feature selection and combination for texture feature extraction.Based on GF-1 WFV satellite image data,this study selects Acheng District,Harbin City,Heilongjiang Province,extracts index features and texture features,combines them to build a multi-feature dataset,and realizes the classification of typical crops in the study area based on various machine learning algorithms,and the classification results are compared and analyzed.The main research contents and results of this paper are as follows.(1)The study takes remote sensing feature extraction and selection as the entry point,vegetation index feature extraction based on image spectral data,texture feature extraction using gray co-occurrence matrix,and introduction of the best index factor method to select multiple texture features extracted to obtain the best texture feature combination.The multi-feature dataset constructed by combining the index features with the best texture features is input into the classification model to realize the classification of typical crops in the study area.The results show that the constructed multi-feature dataset can effectively reflect the characteristics of different crops and achieve good classification results in various classifiers,with the overall classification accuracy reaching more than 90% and the Kappa coefficient reaching more than 0.87.Random forest has the best classification results,with an overall classification accuracy of 93.25% and a Kappa coefficient of 0.9148.(2)The random forest method is mostly based on manual experience in the parameter selection process,which is subjective and difficult to determine the optimal combination of parameters.To address this problem,the adaptive weight particle swarm algorithm is used to optimize the parameter selection process of the random forest.Based on the multi-feature data set,the random forest model optimized by adaptive weight particle swarm algorithm is used to realize the classification of typical crops in the study area.The classification results of the random forest model before and after optimization are compared,and the results show that the overall classification accuracy of the optimized random forest classifier is improved by 1.28% and the kappa coefficient is improved by0.0163.(3)To explore the effect of different scales of texture feature extraction windows on the classification results,the study extracts texture features of different benchmark windows,combines the index features to construct a multi-scale feature dataset,and realizes the classification of typical crops in the study area using a random forest model optimized by adaptive weight particle swarm algorithm.The results show that when the size of the reference window for texture feature extraction is 3*3,the obtained classification accuracy is higher,and with the increase of the size of the texture feature extraction window,the number of irregular patches increases,which makes the overall classification accuracy show a trend of gradual decrease.