Detection Of Public Security Sensitive Targets In Remote And Scattered Areas Based On Multi-source Remote Sensing Data

Detecting public security sensitive targets is an important part of anti-terrorism in Xinjiang.Timely,accurate and efficient acquisition of the distribution of public security sensitive targets in remote and scattered areas is the most important basis for anti-terrorism,stability maintenance and emergency response.Recently,the rapid development of high-resolution remote sensing technology provides high-quality data sources for public security applications,and has become one of the most effective tools for Xinjiang public security organs to obtain spatial intelligence in remote and scattered areas.Compared with data from singe remote sensor,which lacks spectral band information,data from multi-source remote sensors data can provide a more comprehensive understanding of the characteristics of public security sensitive targets,in-depth analysis,and complementary advantages.The research contents and contributions of this paper are as follows:I.Research on public security sensitive target detection based on multi-scale and key features location of high-resolution images(1)Taking advantage of the high spatial resolution of GF-2 panchromatic and multispectral fusion image,considering the size,the shape and the material of the target in remote and scattered areas is diverse,and the distribution between the target and the background category is very uneven,an image semantic segmentation model based on atrous and residual network(ARnet)is proposed.On the basis that the U-shaped network structure can take into account both high-level semantic information and low-level spatial information,we use atrous spatial pyramid pooling(ASPP)scheme to combine multi-scale context information,and atrous convolution is used to expand convolution without increasing computational costs and network parameters.The residual module solves the problem of training difficulties caused by the depth of the network.The test IoU values of ARnet on building and road datasets are 0.8038 and0.8281,respectively,which outperform traditional methods for 2% and 1.78%.(2)From the perspective of key feature positioning and dynamic multi-scale context information extraction,a network of two stream Unet++ for segmentation based on dual attention(TUDA)is further proposed.The unified feature extraction of road and building images can achieve mutual gain between road and building features,and then the dual attention mechanism is introduced into the U-net++ network to achieve key feature positioning.The test IoU of TUDA on the building data set is 0.8749,which is 3.6% higher than other methods,and the test IoU on the road data set is 0.8779,which is 1.9% higher than other methods.II.Research on public security sensitive target detection based on the features enhancing of GF-2 and GF-6 imageTaking advantage of the increased spectral bands based on the fusion image of GF-2panchromatic and GF-6 WFV multispectral image,in order to reduce the impact of the time imsynchronization and the large difference in spatial resolution of the two satellite images and enhance extraction features,a multi-scale pyramid network based on deep feature fusion(MSPN)is proposed.Strip pooling can effectively capture long-distance context information,the scale-aware pyramid fusion module is used to dynamically fuse multi-scale context information in advanced features,and multiple global pyramid guidance modules,which provide different levels of global context information for the decoder,are added between the encoder and decoder.The test IoU of MSPN on the building data set reaches 0.8911,which is1.6% higher than the TUDA network,and about 2.3% higher than the comparison network model.The test IoU on the road data set reaches 0.8828,which is 2.5% higher than the TUDA network.It is about 3.8% higher than the comparison network model.III.Research on public security sensitive target detection based on spectral and spatial features joint sparse representation of hyperspectral multi-source data(1)Taking advantage of the higher spectral resolution of hyperspectral images and assisting with Lidar data,a residual fusion-based representation classifier(RFRC)is proposed.Combining the advantages of global spatial features,local spatial features and spectral features,which avoids the disadvantage point that a single data source cannot fully characterize the target.At the same time,the cooperative mechanism of collaborative representation classifier and the competition mechanism of sparse representation are combined to improve the sparse representation classifier.The IoU of roads reaches 0.8879,and the IoU of buildings reaches 0.8673 in the MFUUFL Gulfport multi-source data for joint detection,a Washington DC data image of a road’s IoU reaches 0.9965,and a building’s IoU reaches 0.9971.(2)Considering the Hughes phenomenon of hyperspectral data requires reasonable feature fusion processing to improve the detectability of the target,this paper then proposes a spatially weighted features fusion based on RCCA(SWFF).While extracting the spectral features,a theory based on collaborative representation is used to calculate the similarity of the spatial neighbor pixels,which can reduce the problem of spectral diversity caused by factors such as sensors and weather.The RCCA(Randomized Nonlinear Canonical Correlation)feature fusion algorithm can retain the discriminable information in sparse and low-rank features to the greatest extent.The IoU of roads reaches 0.9056 and the IoU of buildings reaches 0.8722 in the MFUUFL Gulfport multi-source data for joint detection,the IoU of roads reaches 0.9953 in the Washington DC data image,and the IoU of buildings reaches 0.9961.The research results of this paper provide conditions for the innovation of public security remote sensing applications.Extracting public security sensitive target information based on multi-source data,combining with other public security actual combat auxiliary information aggregation analysis,multi-point joint,scientific deployment and control in actual combat can quickly and effectively block key nodes and roads to achieve efficient emergency response and improve the efficiency of anti-terrorism handling.