Research On Feature Representation For Remote Sensing Data Analysis

The feature representation of remote sensing data is an essential premise for information extraction and various applications.This thesis proposes some feature representation methods to address the high dimensionality,small sample size,nonlinear distribution,multi-source information fusion and noise corruption issues based on the theory of machine learning and pattern recognition.To validate the effectiveness of these methods,we apply the learned representation to retrieve atmospheric aerosol optical depth(AOD)and classify remote sensing images.The main contents of this thesis are summarized as follows:(1)Graph-regularized low-rank feature representation(GLRR)for AOD retrieval.Remote sensing data is easily corrupted by various noises in the acquisition process.Besides,many works have concluded that the high-dimensional remote sensing data often lies in a low rank subspace.Therefore,we firstly adopt a low-rank representation(LRR)model to learn a powerful representation from the corrupted spectral response.Then,a graph regularizer is incorporated into the LRR model to capture the local structure information and the nonlinear property of the remote sensing data.Since it is easy to acquire the rich AOD results retrieved by satellites,we use them as a baseline to construct the graph.To optimize the GLRR model,we adopt a linearized Alternating Direction Method of Multipliers(ADMM)algorithm,dividing it into three sub-problems and iteratively optimizing them one by one.Finally,the learned feature representation is fed into Support Vector Machines(SVMs)to retrieve AOD.(2)Ridge regression(RR)based AOD retrieval.This thesis proposes two AOD retrieval algorithms based on the RR model.The first one combines the spectral band selection and AOD retrieval processes into a RR model,resulting in an optimal band subset for the AOD retrieval.Meanwhile,a semisupervised method is used to make full use of the information from unlabeled samples.Based on these two ideas,we can effectively address the issue of using small numbers of training samples with high-dimensional featuers.Besides,to capture the nonlinear distribution property of remote sensing data,this thesis uses a kernel method to map the original data into a high-dimensional Hilbert space,where a linear regression algorithm can retrieval AOD.The second algorithm attempts to fuse the advantages of the physical model and the machine learning model,and proposes two hybrid frameworks based on RR to correct the retrieval bias.In the serial framework,the retrieved AOD by the physical model is used as a feature for RR.In the parallel framework,RR is directly used to learn the retrieval bias between the estimated AOD by the physical model and the groundtruth AOD.(3)Matrix discriminative analysis(MDA)model for hyperspectral image classification.This thesis proposes two MDA based spectral-spatial feature learning models.For the first model,a matrix-based spectral-spatial feature representation method is designed for each pixel in the hyperspectral imagery to capture the local spatial contextual and the spectral information of all bands.Then,MDA is adopted to learn the discriminative feature subspace for classification.To further improve the performance of the discriminative subspace,a random sampling technique is used to produce a subspace ensemble for the final classification.Besides,due to sensor interferers,calibration errors,and other issues,hyperspectral images can be noisy.These corrupted data easily degrade the performance of MDA.Therefore,we farther propose a second model named robust MDA(RMDA)to address this important issue.Specifically,based on the prior knowledge that the pixels in a small spatial neighborhood lie in a low-rank subspace,a denoising model is first employed to recover the intrinsic components from the noisy imagery.Then,MDA is used to extract discriminative spectral-spatial features from the recovered components.Besides,different hyperspectral images exhibit different spatial contextual structures,and even a single imagery may contain both large and small homogeneous regions simultaneously.To sufficiently describe these multiscale spatial structures,a multiscale RMDA model is proposed.(4)Deep feature learning based high-resolution remote sensing scene image classification.This thesis proposes two SPP-net based deep learning models to learn the rich texture features of high-resolution remote sensing scene images.The first one is a multi-scale deep feature learning model.First of all,the original scene image is warped into multiple different scales.Then,these multi-scale images are fed into the ImageNet pretrained SPP-nets to extract multi-scale deep features.Finally,a multiple kernel learning method is developed to automatically learn the optimal combination of these features.The second one is a new adaptive deep pyramid matching(ADPM)model.Inspired by the spatial pyramid matching(SPM)model,we consider features in all convolutional layers as a multi-resolution representation of the input image.Then,the pyramid matching kernel is used to combine them into a unified representation.Different from SPM,we replace the low-level descriptors as deep features,and the optimal fusing weights among different convolutional layers are learned from data itself,instead of pre-defined values.Besides,to reduce the information loss caused by a fixed size input image,multi-scale images are fed into the SPP-net to learn complementary information from different scales.