Remote Sensing Image Primary Object Extraction Based On Image Representation Learning

The remote sensing images obtained through remote sensing technology provide convenience for researchers to explore the world.The remote sensing images acquired in different ways contain different resolutions,spectral bands,coverage aera and other information.Remote sensing images can monitor the changes of surface objects in real time from a highaltitude perspective,which can be used in fields such as mapping survey,urban planning,and resource detection.However,there are still some problems in the field of remote sensing image feature extraction,such as the remote sensing image itself is diverse,and different network structures are designed according to different image characteristics.The complex imaging mechanism of remote sensing images may have redundant information,which requires the selection of key information from a large amount of information for image representation.In addition,remote sensing images contain complex content,so it is necessary to consider useful information from different perspectives when designing classifiers.Based on these issues,this paper makes an in-depth study on the primary object extraction of remote sensing images from the two directions of hyperspectral image classification and aerial image building extraction.The main research contents are as follows:1.A hyperspectral image classification method based on end-to-end multi-scale deep learning network is proposed.Combined with the idea of segmentation network,an end-to-end network based on residual module is proposed for HSI classification task,which is used to extract spatial information in HSI.The method also keeps the number of channels in the network structure constant to extract and retain the spectral information in HSI.In addition,for the complex objects with various sizes in HSI,this chapter also adopts a simple and effective multi-scale module to extract multi-scale information.The results on three public data sets show that the model in this chapter can outperform the comparison algorithms on the HSI classification tasks.At the same time,this method does not require the preprocessing step of cropping pixel blocks,which reduces hardware consumption and classification time.2.A hyperspectral image classification method based on interactive attention mechanism deep learning network is proposed.The method in this chapter introduces a channel attention mechanism to reduce the redundancy of spectral information in HSI.Based on the channel attention mechanism,a three-branch interactive attention mechanism is designed to extract the interactive relationship between dimensions in HSI.In this chapter,the interactive attention mechanism network is used as the backbone network,the residual block is used as the auxiliary structure,and the results of the three branches are fused as the final classification result.The classification results on three public data sets show that the model in this chapter can extract effective spatial-spectral features,and can improve the accuracy and efficiency of the HSI classification task.3.An aerial image building extraction method based on global and multi-scale encodingdecoding network is proposed.Aiming at the characteristics of aerial images,this method uses an encoder-decoder network with skip connection as the main frame.In the encoder,VGG16 is used to extract local information,non-local block is used to obtain global information,and local and global information are better integrated through the connection block.The decoder extracts basic information and multi-scale information from aerial images through deconvolution branch and multi-scale branch respectively to improve the distinguishability of image features.The results on two public building data sets prove that the method in this chapter can obtain good extraction results on the task of building extraction.