Cloud Processing Methods For Optical Remote Sensing Mapping Of Urban Green Spaces

As one of the important components of urban space,urban green spaces(UGS),such as urban green lands,urban wetlands,and urban forests,perform an important role on alleviating the urban heat island effect,regulating urban climate,serving residents’ leisure and so on,providing essential data for the urban ecosystems,urban carbon sinks,urban sustainable development and related studies.With the advantages of wide coverage,high observation frequency and complete spectral information,optical satellite remote sensing images make it possible to obtain urban green space information quickly.However,clouds are inevitable problem in optical satellite,especially in tropical and subtropical regions,which can directly lead to the discontinuity of optical remote sensing images in spatial patterns,and meanwhile bring uncertainty to UGS mapping based on optical remote sensing images.Aiming at the specific problems of cloud processing in the current UGS mapping based on optical remote sensing images,the optical remote sensing images "cloud detection-cloud removal-information extraction" are taken as the main research line,the improvement of the utilization rate of optical remote sensing images and the extraction accuracy of UGS mapping is taken as the research purpose in this thesis.The cloud processing method has been researched for optical remote sensing mapping of UGS,and a series of reliable cloud processing(cloud detection and cloud removal)and UGS extraction algorithms have been developed,which includes:(1)In terms of cloud detection,given that the current methods face false detection of bright surfaces and clouds and the limitation of spectral range,a deep learning cloud detection method by integrating public geographic information has been proposed.Firstly,a public geographic information encoder was constructed to convert public geographic data as auxiliary maps.Secondly,multi-scale convolution spatial-spectral and geographic information module was designed to extract the spatial-spectra-contextual semantic features of clouds from remote sensing images and public geographic data,aiming to distinguish clouds from background surfaces.Finally,the Landsat-8 satellite,Sentinel-2satellite,and Gaofen-1 satellite are tested on a global scale,with the results showing that the proposed method can obtain high accuracy of cloud detection for bright surface areas.(2)In terms of cloud removal,a novel spatiotemporal deep neural networks(ST-net)was proposed to address the problem of the spectral distortion and blurred spatial details after reconstruction in multi-temporal methods.The ST-net consists of spatial-temporal learning module,spatial-temporal feature fusion module,and reconstruction module.It makes full use of the temporal-spatial-spectral features of multi-temporal remote sensing images,which generate cloud-free images in terms of temporal-spatial-spectral.The cloud removal experiments were conducted using Landsat images and Sentinel-2 images as experimental data.The results show that the ST-net method can generate spectrally high fidelity and spatially seamless cloud-free images so as to improve the quality of subsequent spatiotemporal continuous UGS mapping.(3)In terms of UGS mapping,given that a large number of labeled samples in deep learning-based UGS mapping method are hard to obtain,an automatic generation method of high-fidelity samples based on crowdsourced geospatial data was studied to solve the problem.On this basis,a UGS mapping method based on feature adaptive pooling CNN was proposed to solve the problem of missing extraction of UGS.A feature-adaptive pooling CNN was designed to extract the discriminative features of UGS.The UGS extraction experiments were conducted for Sentinel-2 images in the case of a small number of labeled samples.Results show that the proposed method can provide reliable UGS mapping results and effectively suppress the missing extraction problem of UGS and background surface.(4)In terms of prototype system development,based on the optical satellite cloud processing method and UGS extraction method,an optical satellite oriented cloud processing and UGS intelligent acquisition system was developed.Firstly,producing remote sensing products of spatiotemporal continuous UGS is set to the target,the overall architecture of the prototype system is constructed,and then the prototype system is developed.Finally,cloud detection,cloud removal and UGS extraction module of the prototype system are tested.The results show that the prototype system can obtain high accuracy of UGS maps.The practicability and reliability of the prototype system are preliminarily verified for producing spatially continuous UGS maps.