Target Recognition In Urban Remote Sensing Images Based On Multi-feature Space And Its Optimization

At present,remote sensing technology is becoming more and more developed,which has become the technical support of many scientific research and industrial applications.At the same time,it has also produced new research directions.Remote sensing image target recognition benefits from the continuous progress of remote sensing technology,and with its own sustainable development and wide application,it has gradually become a research hotspot.Among them,urban remote sensing image target recognition can distinguish the categories of urban features,which is the best means of urban environmental monitoring.However,the pixel-based methods can not make full use of the rich feature information of high-resolution remote sensing images,and the object-based methods can not obtain accurate objects by existing segmentation technology.Both of them have different limitations.Aiming at the shortcomings of traditional methods,this paper proposes a target recognition method of urban remote sensing image based on multi-feature space and its optimization.This method combines pixel features,object features and depth features to construct a huge multi-feature space,optimize the multi-feature space,and finally send it to the classifier,so as to achieve the purpose of target recognition of urban remote sensing image.The specific research contents include:(1)53 object features are extracted by using the multi-scale image segmentation algorithm,34 pixel features are extracted by using the chessboard segmentation algorithm,36 depth features are extracted after the convolution layer of VGG19 network,and a multi-feature space containing 123 features is constructed.(2)Optimize the huge multi-feature space,comprehensively evaluate the four commonly used feature importance algorithms according to the standard of obtaining high recognition accuracy with a small number of features,adopt XGBoost algorithm to sort the feature importance,and finally find that the combination of the top 18 features in importance can obtain high recognition accuracy.Therefore,they form an optimal multi-feature space.(3)After successfully constructing the random forest recognizer in the best state after parameter adjustment,in order to explore the accuracy of this method,the traditional method and this method are applied to the study area respectively;In order to explore the effectiveness of this method,this method is applied to the verification area.The experimental results show that the overall accuracy of the same research area based on pixel,object,multi-feature space and its optimization method is 77.28%,82.45% and 87.89% respectively,and the overall accuracy of the verification area based on the application of this method is 85.39%,which proves that this method has certain accuracy and effectiveness,and is a feasible method of urban remote sensing image target recognition.