Dimension Reduction Of Remotely Sensed Data

In remote sensing applications,data obtained from a single sensor often fail to meet certain needs.The integration of datasets from different sensors can improve the applicability and recognition ability of ground objects on the Earth's surface.Compared with the remotely sensed data obtained from a single source,the fused data usually have the characteristics of redundancy.There is usually a strong correlation among different bands.When the multi-source remotely sensed datasets are applied,the redundancy can cause huge computational demand.Therefore,one needs to reduce the dimension of the datasets to minimize the comsumption of computation time and storage space.Thus,the objective of this thesis is to extract the intrinsic low-dimensional structure of highdimensional data through a manifold learning approach.T he main content includes:(1)For the problem that the traditional manifold learning algorithm cannot deal with the large-scale dataset,this thesis constructs a complete framework for the reduction of the multi-source dataset.The intrinsic dimension estimation is performed on the dataset to determine the target dimensionality.Then,a subset called as the landmarks is sampled from the dataset.Next,the dimension reduction is performed on the landmarks w ith the manifold learning approaches to obtain the manifold skeleton.Finally,the unsampled points are inserted into the manifold skeleton.In the evaluation of the overall performance of the framework,the obtained data are classified with the random forest(RF)and support vector machine(SVM)classifiers.(2)Some manifold learning algorithms such as isometric mapping(Isomap)and local linear embedding(LLE)are sensitive to the number of neighborhood.This thesis utilises L1 norm to construct he graph of neighborhood,and our proposed algorithm can adaptively select the number of neighborhood.(3)The framework is tested with the five nonlinear manifold learning algorithms including isometric mapping(Isomap),diffusion map(DM),t-distribution stochastic neighbor embedding(t-SNE),Gaussian process latent variable model(GPLVM),and the improved Isomap(L1-Isomap)for dimension reduction using the TSX,RS2,and ALOS data.At the same time,the random forest(RF)and support vector machine(SVM)classifiers are selected to evaluate our proposed method fairly.The experimental results showed that 1)the improved Isomap(L1-Isomap)achieved the highest over accuracys(OAs)among the five nonlinear manifold learning methods with RF and SVM classifiers for the TSX,RS2,and ALOS data;and 2)the OAs yielded by L1-Isomap was comparable or even better than those of the raw datasets.Therefore,the improved Isomap(L1-Isomap)is robust in dimension reduction and classification for multi-source remote sensing data.