Restoration And Classification Of Remote Sensing Imagery Based On Multiple Feature Learning

Due to sensor malfunction and poor atmosphere,the acquired optical remote sensing images are incomplete.The existing missing characteristics hinders the follow-up interpretation.With the development of different spectral,spatial and temporal resolution for earth observation technologies,it requires new multisource data fusion methods to associate spatio-spectral-temporal features and restore a high resolution image.In recent years,hyperspectral image analysis has attracted more attention on the community of remote sensing.However,the ratio between the labeled training samples and the number of spectral bands is low,it is desirable to develop new multiple feature learning models to increase the classification accuracy with limited training samples.Moreover,the research on remote sensing data and social media data fusion provides a new way to handle limited training samples problem.We can utilize the ground street view data to obtain the semantic information behind the scene,which could guide the process of aerial image semantic segmentation.This dissertation deals with the issues of remote sensing image restoration and classification.Firstly,we work on optical image missing characteristics restoration with multitemporal data.Then,we propose a new model to handle multisource data fusion.After that,we use multiple feature learning method to increase the ability of hyperspectral interpretation.Finally,we present a novel cross-view domain adaptation model to fuse ground streetview data and remote sensing images and increase the accuracy of weakly supervised semantic segmentation.This dissertation has the following characteristics and technical contributions:1.A coupled neighborhood regression model is proposed for remote sensing image missing characteristics restoration.We first utilize cluster method and sparse representation to learn the category specific mapping relation between missing information and real information.After that we learn the nearest neighbor of each subdictionary's atom and use these features to efficiently restore the missing characteristics.In order to associate the temporal feature difference caused by geographical change,we propose a new common feature space learning model,which increases the restoration accuracy.The experimental results with spectral and spatial missing characteristics have showed that the proposed method could learn several useful mapping functions with the common feature space.The proposed method could efficiently and effectively increase the restoration accuracy and deal with the problems caused by long-term geographical changes,which could be easily extended to real world applications.2.A new integrated spatio-spectral-temporal sparse representation model is proposed to deal with multisource data fusion.Previous research on multisource data fusion is limited to specific sensor type and few efforts have been made to exploit the relationship among heterogenous data sources with different resolutions.Motivated by previous research on human vision system,we propose a spatio-spectral dictionary learning method,which maintains the high spectral correlation and high spatial self-similarity of different data types.Then,we use a local constraint weight learning model to associate the temporal geographical changes.We introduce an analytical solution and the alternating direction method of multipliers(ADMM)optimization methods to lower the computational complexity.The spatio-spectral,spatio-temporal and spatio-spectral-temporal experimental results illustrate that the proposed method based on the learned spatio-spectral-temporal features could provide high resolution data to address demanding work in real-world applications(i.e.natural disasters).3.A novel multiple feature learning method based hyperspectral image classification method is proposed to solve the limited training samples problem.We first present a novel similarity kernel and diversity kernel to keep the spectral similarity and spatial neighborhood correlations.Then we introduce a kernelized Bayesian feature projection method to choose the most useful features for each classifier.The weights learning part guarantees the specific class weights for each feature type.Finally,we propose a decision fusion method to associate the linear and nonlinear features and improve the classification performance.The experimental results using three hyperspectral images show that the proposed method could outperform most of the state-of-the-art methods even with limited training samples.Moreover,the proposed method is extended to fuse hyperspectral and LiDAR data to demonstrate its performance in multisource data fusion applications.4.A new aerial and ground street-view data fusion method is proposed for crossview aerial image semantic segmentation.The process of remote sensing image annotation often costs a lot.So we utilize a new domain adaptation method,which use intra-domain and cross-domain adversarial loss to lean the shared and private feature maps.Then we use the fully convolutional networks to extract the exact semantic information of remote sensing images.The introduction of output adversarial loss maintains the cross view spatial similarity of the semantic results.The experimental results with the popular aerial semantic segmentation competition dataset illustrate that the proposed method could utilize the annotations from the street-view data to pursue the scene distribution of the aerial image and improve the semantic segmentation performance.Moreover,we design a new experiment that uses the aerial image to deal with street-view images orientation estimation and fine-grained geocalibration,which demonstrate that its application is feasible and promising.In conclusion,the dissertation studies multitemporal remote sensing image restoration,multisource data fusion with different resolutions,hyperspectral image classification,remote sensing and street-view data fusion and geocalibration.This research solves the issues from remote sensing data acquisition to image interpretation and semantic segmentation with limited training samples.It provides technical support for remote sensing applications,which shows its theoretically and practically promising future.