Topic Model For High Resolution Remote Sensing Imagery Semantic Scene Understanding

With the rapid development of high spatial resolution(HSR)remote sensing sensor technology,the resolution of HSR images obtained gradually increases.How to effectively utilize the rich spectral and spatial details for accurate image interpretation has become the most challenging research front in the field of remote sensing.With the deep study of HSR image information extraction technology,the interpretation of HSR images basically realizes the transformation from pixel-oriented classification to object-oriented classification.This enhances the object classification performance.However,current ground object classification is achieved mainly by extracting the low-level features from the images,and can only reach the ground object level,such as trees and buildings.It is difficult to obtain the high-level scene semantics,such as residential areas and commercial areas.To acquire the high-level semantics,it has become a core and hot issue to cross the gap between the low-level information and high-level semantics in the HSR imagery.Rooted in natural information processing,the use of the probabilistic topic model(PTM)to capture latent topics to represent HSR images has been an effective way to bridge the semantic gap.However,how to effectively discover discriminative information to recognize the complex HSR scenes,the following issues are encountered:1)Limited ability to learn the low-level features.In general,a single feature is utilized to describe the HSR scenes.Methods combing multiple features usually simply concatenated the different features,which ignore the feature difference.This leads to limited remote sensing scene descriptive ability.2)High redundancy and limited sampling method of mid-level features.The generated topics of the classical PTM are dense and redundant,which leads to high time consuming and low discrimination of HSR scenes.Heterogeneous descriptions can be obtained for the uniform grid sampling based scene classification methods.However,it does not perform well for HSR scenes with representative objects.3)Limited ability to understand high-level semantics.Existed scene classification method ignore the spatial position information between the image patches and the detailed global description of HSR scenes,ans is thus difficult to understand the high-level semantics in the HSR scenes.To overcome the challenges of topic model for HSR scene interpretation,this thesis develops scene classification methods based on PTM model for HSR images from three aspects of low-level feature description,mid-level topic modeling,and high-level semantic understanding.The major works and contributions of this thesis are listed below:(1)The characteristics of HSR image scenes,the current state and the problems of HSR image scene understanding are systematically summarized in this thesis.The basic theory of PTM is introduced,and this thesis also specifically analyzes the advantages and practical potential of PTM for HSR scene understanding.(2)In the aspect of low-level feature description,multi-feature fusion PTM for HSR image scene understanding is proposed.Because of the diversity of the objects and the variability of the distribution,the heterogeneous features are captured from the local and global,discrete and continuous perspective in this thesis,to improve the scene descriptive ability of bag-of-visual-words(BOVW)model.In addition,the traditional methods simply concatenate the multiple features during dictionary construction and topic modeling,which ignore the problem of feature difference.In this thesis,the spectral,texture,and SIFT feature are separately extracted for topic modeling and are fused at the topic level,to improve the scene interpretation performance.(3)In the aspect of mid-level topic modeling,sparse homogeneous-heterogeneous topic feature model for HSR image scene understanding is proposed.As the topics mined by the classical PTM are redundant,the Dirichlet distribution in the classical PTM is replaced by the sparse inference approach,and the fully sparse semantic topic model is proposed.In addition,based on the union of SLIC superpixel sampling and uniform grid sampling,this thesis exploits both the heterogeneous and homogeneous information,to decrease the redundancy of the topics and speeds up the scene understanding for HSR images.(4)In the aspect of high-level semantic understanding,deep sparse semantic modeling framework for HSR image scene understanding is proposed.In general,the BoVW model disregards the spatial layout information,whereas deep learning can preserve the spatial location,but it may lose the characteristic information of the HSR images.In this thesis,based on the BoVW model and deep learning,the features are extracted from three aspects of low,middle,and high level,to effectively construct the multi-level description of complex scenes.In addition,by analyzing the distinct characteristics of deep learning and PTM,an adaptive deep sparse semantic modeling framework is proposed to achieve the accurate interpretation of complex scenes.(5)HSR image scene understanding framework based on PTM is built.Combined with HSR image scene understanding method based on PTM proposed from multiple perspectives,the HSR image scene understanding framework based on PTM is built for different demands.This thesis conducts the research of PTM in low-level feature description,mid-level topic modeling,and high-level semantic understanding for HSR image scene understanding.This is able to improve the performance of HSR scene understanding,and further promotes the practical application of HSR scene understanding,therefor have significant scientific and social values for many fields,such as change analysis of urban functional area and environmental monitoring.