The Extraction Of Typical Geographic Features For Domestic High-resolution Satellite Remote Sensing Images

In recent years,with new demands were continued to put forward for natural resources and urban ecological environment monitoring,extracting typical features,such as coastlines,farmland,buildings and roads from remote sensing images has become a hot topic of industry and academia.However,traditional remote sensing images have low resolution so that it is difficult to acquire fine-grained scale objects information from them.Moreover,the previous studies were mostly rely on foreign remote sensing images.With the international situation changes,it is easy to face the "stuck neck" problems.With the rapid development of domestic high-resolution satellites remote sensing(in short GF)technology,batches of GF satellites,such as ZY-3,GF-2,and GF-3,have been successively emission.High-resolution remote sensing images can obtain rich detailed information on the surface,providing data support for the fine extraction of nature resource elements.Using GF images to extract typical ground objects will provide decision-making support for urban planning,agricultural resource monitoring,ecological environment assessment and other fields.This has become a current research hotspot.However,objects show more complex characteristic in the high-resolution remote sensing images.For example,urban buildings and roads have complex structures,diverse types and large scale variation.Traditional machine learning algorithms mainly extract middle and low-level features based on manual methods,which have shortcomings in feature expression and obvious limitation in classification accuracy.Tradition coastline extraction are mostly based on binary classification methods,which are lower intelligent and require more post-processing.Farmland presents extremely complex characteristics in terms of spectral features,spatial features and temporal features,which impose higher requirements on extraction algorithms.With the higher requirement in accuracy and efficiency of extraction algorithm in production practice,it is an inevitable trend to develop intelligent typical objects extraction algorithms.The "Guidelines for development of scientific and technological innovation in Natural Resources " issued by the Ministry of Natural Resources in 2018 also clearly pointed out that it is necessary to strengthen the ability to fast extract and intelligently interpret full factors information of natural resource based on the multi-source investigation and monitoring results.This paper focus on the problems faced by the extraction of typical objects from domestic highresolution remote sensing images,especially four core features such as coastlines,farmland,buildings and roads.Based on the in-depth study of traditional machine learning algorithms and combining deep learning networks,we develop an intelligent extraction method of feature elements suitable for domestic highresolution remote sensing images.The main research contents of this paper are as follows:1.This paper proposed a novel method for coastline extraction from GF-3 SAR images.In the method,based on the Gaussian Mixture Model(GMM),a new K-distributed local statistical active contour model(LKDACM)is designed to complete the identification and extraction of the coastline.At first,the sea-land was segmented roughly by GMM to reduce the number of iterations of the subsequent model.And then the details of the complex area in SAR images were described by LKDACM.Results show that the proposed coastline extraction method overcomes the shortcoming of the traditional method and extracts coastline fast and accurately from SAR images.2.Aiming at the problem of farmland extraction from high-resolution remote sensing images,this paper proposed a pixel-level method based on the improved SAM model.At first,in order to reduce the difficulty of sample preparation,the method uses seed points for sample selection.Subsequently,the fast image texture analysis algorithm was used to obtain texture features and the normalized vegetation index was introduced to improve the separability of farmland and others.Finally,the farmland is extracted based on the adaptive threshold K-Means algorithm.The domestic high-resolution remote sensing images,such as GF-1 WFV,GF-2,ZY1-02 C were used to verify the capability of new method to extract farmland.The results show that the new method has stronger adaptive ability,higher accuracy than traditional method.And the robust extraction effect can be obtained regardless of in plains or mountains,even in different seasons.3.In response to the building extraction from high resolution remote sensing images,this paper proposed a semantic segmentation network based on improved U-Net model.In order to acquire more high frequency information,the new network upsampling the input images at first.And then it reduces the number of channel features for alleviate the hardware memory pressure.Apart from this,in order to make better use of mid-level features in the model,the new network resizes the difference feature maps in the decoder as the same as input images.And the final segmentation results comprehensively consider features of different depths.The new network has achieved higher accuracy than classic semantic segmentation networks such as FCN,SegNet,U-Net,PSPnet through testing on the public Massachusetts datasets.A dataset was made from GF-2 to verify the classification effect on domestic images,the new network also achieved the best results among all the networks mentioned above.4.There are problems such as imperfect roads edge and poor topological connectivity in extracting roads from high resolution images by semantic segmentation model.In terms of this issue,this paper proposed a multi-scale semantic segmentation model which based on U-Net and enhanced the roads edge information.The model added an edge extraction network in parallel to the original road segmentation network by integrating features maps of different levels.In the result,learning for roads edge was strengthen.Furthermore,the network adds a cascaded dilated convolution model after the deepest feature map so that both local details and global information can be considered.The experimental results on the public Massachusetts dataset show that the model in this paper can achieve better classification results than SegNet and U-Net.In addition,this model also achieved better classification results than other networks in the dataset from GF-2 images.The result indicates that the proposed model can extract roads in domestic satellite images with high accuracy.