Research On Winter Wheat Extraction Method Based On Semantic Features And Statistical Features

Accurate information on the spatial distribution of crops is of great significance for food production estimation,agricultural policy making,and scientific research.At present,the convolutional neural network has become one of the main methods for extracting the spatial distribution information of crops from remote sensing images,but by applying the convolutional neutral network solely,there will be some disadvantages such as rough edges.Unreasonable classification of crops at the edge of the crop growing area is a major problem in image segmentation using convolutional neural network.Improving the accuracy of edge pixel classification is the key to obtain high-precision spatial distribution information of crops from remote sensing images by using convolutional neural network.In response to the need to extract high-precision crop spatial distribution information,this thesis selects winter wheat as the research object,and on the basis of fully analyzing the relationship between the structure of the convolutional neural network and the extraction results,taking account of the respective advantages of semantic features and statistical features,it proposes a step-by-step method for extracting spatial distribution information of winter wheat.Firstly,the RefineNet model is improved,and the improved RefineNet model is used for initial extraction;Secondly,on the basis of the analysis of the initial extraction results,the confidence coefficient is used to select the pixels that need to be further classified in the initial extraction results.The semantic features and statistical features are selected to establish the feature vector,and the deep belief network(DBN)is used to extract the precise spatial distribution information of winter wheat.The specific research contents are as follows:1.Use remote sensing image processing software to perform preprocessing such as radiation calibration,atmospheric correction,orthorectification,and image fusion on 37 GF-2 remote sensing images covering the winter wheat area of Taian from 2018 to 2019.Then the geographic information system software is used to mark the real category of three selected pre-processed images,and the original images and the marked images are made into the data set of the thesis experiment.2.Improvement on the traditional RefineNet model mainly includes:(1)Improvement of feature fusion method.Make access to 1*1 convolution layer after upsample to integrate channel information;Increase the proportion of low-level high-resolution feature maps in the fusion process;(2)Improvement of classifier.Improve the output part of the Softmax layer to output the predicted category and Confidence value of each pixel.The improved RefineNet model is used to extract the initial data,and the test results are statistically analyzed.The pixels that need to be extracted are selected through the confidence coefficient.3.Based on the similarity in color and texture of winter wheat pixels,the Euclidean distance can be adopted to analyze the similarity of vectors.In order to make full use of the advantages of semantic features and statistical features,in fine extraction,the semantic features generated in the initial extraction process,texture features of the original image,texture statistics features and Euclidean distance features are applied to form a feature vector to make a comprehensive description of the pixels to be extracted.DBN can effectively mine the essence and deep features of the data,therefore it is used to train the combined feature vectors and the real categories of pixels to obtain a weight file,and then this weight can be used to achieve fine extraction and reclassification of untrusted pixels after initial extraction.This thesis selects SegNet and original RefineNet as the comparison model,and designs the comparison experiment.The experimental results show that the precision rate(93.4%),recall rate(94.1%)and accuracy rate(95.1%)of the method in this thesis are obviously better than those of the contrast model.It is proved that the proposed method has certain advantages of significantly improving the recognition results of crop edge regions and improving the recognition accuracy in remote sensing image segmentation.