Morphology Feature Extraction And Its Multi-level Classification Based On Pixel And Scene For High Resolution Remotely Sensed Imagery

With the progress of urbanization in China,population distribution,social activities,and energy consumption have been highly concentrated in urban areas,which are having an increasingly strong impact on the Earth's environment.Under this background,the development of eco-city is favorable for the goal of ecological environmental protection,people's pleasant living situation,and social sustainable development.Therefore,how to monitor urban environmental elements has drawn much attention from domestic and overseas researchers.In pace with the rapid development of sensor technology,high-resolution remote sensing images can provide urban environment study with multi-level essential data.Among them,land covers and urban functional zones are two kinds of basic information.However,there are a series of problems and challenges for the accurate information interpretation of high-resolution remote sensing(HRRS)imagery.Firstly,due to the limitation of the satellite imaging technology,the improvement of spatial resolution restricts the spectral resolution of HRRS images,which decreases the spectral separability of different land cover classes;secondly,it is difficult for the classification of HRRS images to delineate the detailed and structural information,as well as the multiscale features of objects;lastly,there is a semantic gap between low-level data and high-level scenes for HRRS.To address the above issues,based on the morphology features of geospatial objects,we focus on feature extraction,land cover classification,and urban functional zone identification for HRRS imagery.In this way,multi-level classification algorithms in terms of feature,label,and scene can be developed.Afterwards,environmental noise analysis in Shenzhen,as a case study,is conducted for validation and application of the proposed algorithms.Finally,the main contents and corresponding contributions of this thesis are concluded as follows:(1)We summarize the basic methods of morphology feature extraction and classification as a review.At first,in the part of morphology feature calculation,mathematical morphology based on the spectral space,and landscape metrics based on the class space are introduced in detail.Then,for the classification based on morphology features of geospatial objects,two kinds of processing units,i.e.,pixel and scene,as well as the machine learning approaches are presented.(2)Generalized differential morphological profiles(GDMPs)are proposed for spatial/structural feature extraction,aiming to complement the spectral signature of HRRS images.By measuring the morphological difference between arbitrary scales,GDMPs can obtain complete shape spectrum,and effectively describe multiscale characteristics and complex urban landscapes.Moreover,the proposed GDMPs are investigated under circumstances of both geodesic and partial reconstruction.The experimental results have shown that,compared to the traditional DMPs,GDMPs are more appropriate for the enhancement of spectral separability,and hence show a better classification performance.(3)The edge-preservation multi-classifier relearning framework(EMRF)is proposed for land-cover classification.To better characterize spatial distribution of land cover,this method proposes relearning module based on landscape features(relearning-landscape),which can quantify both landscape composition and configuration for spatial feature representation of the class(labeling)space.Furthermore,in order to preserve detailed edges and structures,we introduce tri-training strategy into relearning-landscape,which reduces boundary uncertainty by mutual learning of multi-classifier system.In the EMRF method,the relearning-landscape and tri-training modules work in a collaborative manner to iteratively optimize classification results from two aspect-features and samples.Relearning-landscape can greatly increase the separability for spectrally similar classes,and hence help tri-training select reliable and informative samples.Subsequently,by courtesy of these newly selected samples,the classification model can be further improved,which facilitate a more accurate classification map and,therefore,landscape features can be updated.(4)Deep convolutional neural network based on morphological features(DCNN-MF)is proposed for scene classification of urban functional zones.In order to better characterize image scenes with the limited number of training samples,we construct networks by fusing transferred pre-trained deep convolutional neural networks(CNNs)and remote-sensing based small CNNs,achieving the deep feature representation of morphological profiles and multispectral information.Furthermore,DCNN-MF is used for the classification of multi-temporal HRRS images in Shenzhen,and the long-term land use change can be quantified.Subsequently,by linking land use classes to ecosystem services(ES),the application of the proposed algorithm for ES monitoring can be analyzed.(5)Lastly,based on the multi-level classification algorithms developed in this paper,we take environmental noise analysis in Shenzhen as a case study,and test their applications for urban acoustic environment monitoring.In details,by classifying HRRS images into land cover and urban functional zoning maps,we can depict urban morphology parameters over a relatively large area for the subsequent analysis with environmental noise.Then,for the first time,the relationships between them can be revealed at the city level.Experiments have validated the effectiveness of the proposed classification algorithms as well as their potential for the study on urban environment.