Research On Key Technologies Of Automatic Mapping Based On Remote Sensing Image Target Recognition And Feature Extraction

Optical remote sensing image is usually composed of three kinds of resolution information: temporal resolution,spatial resolution and spectral resolution.The above three kinds of resolution information represent all kinds of information of optical remote sensing image.Due to the complexity of data and model,conventional remote sensing image processing methods are difficult to deal with the new challenges brought by the complexity of the problem.Aiming at the automatic mapping of remote sensing images,this study focuses on the problem of spectral feature curve without considering pixels in the traditional feature extraction algorithm in the field of remote sensing images,and focuses on the problem of losing spectral features in the dimensionality reduction of hyperspectral images.This study also studies the target detection technology for remote sensing images.The research starts with the existing target detection technology based on convolutional neural network,considers the over fitting phenomenon of the existing model in the target detection task of remote sensing image and the low efficiency of small target detection,and focuses on the optimization strategy of fast r-cnn model.In the study of remote sensing image automatic mapping technology,different scenes of data availability are investigated,the input spatial data obtained from satellite images and digital elevation models are evaluated,and compared with the accurate geodetic data generated from field observations to generate 3D urban models.The specific research work of this study is as follows:1.At the beginning of this study,aiming at the shortcomings of the existing feature extraction methods in hyperspectral images,the constrained particle swarm optimization algorithm is added to the spectral segmentation integration method to globally identify the optimal spectral channel as some parts of the spectral feature curve.Finally,the existing features in each channel are integrated by weighted average operator to extract new features.Compared with the traditional feature extraction methods,the experimental results show that the spectral segmentation integral algorithm optimized by particle swarm optimization can not only reduce the high dimension of the image,but also increase the class separability.It can perform better than other methods in classification accuracy.2.In view of the over fitting problem and low efficiency of the existing methods in small target detection,this study uses an improved method based on trainable activation function to improve the existing fast r-cnn on the basis of the existing fast r-cnn.In the cascade step,the loss value is used to improve the learning process of the object.Fourier series and linear activation method have the advantages of better convergence,which is helpful to solve the over fitting problem.The results show that the improved fast r-cnn model used in this chapter is more efficient.The target detection performance of remote sensing images has also been improved.3.This study uses the deep learning algorithm to automatically extract the features of the research area instead of the traditional technology.Different technologies for generating and analyzing 3D city models based on GIS spatial data and high-resolution satellite images are used.Using 2D digital map and digital elevation information,a realistic 3D model is created for Qingshuihe campus of University of Electronic Science and technology of China.