Multifeature Hyperspectral Remote Sensing Image Classification With Small Sample Size

Due to the limitations of imaging environment and sensors,spectral confusion is a universal phenomenon that the same ground objects have different spectrums while different ground objects have the same spectrum.These facts increase the difficulty and cost of annotating samples.In real scenarios,it is very difficult to provide sufficient training samples to supervised classifiers,which will lead to overfitting.Furthermore,spectral confusion limits the discriminaty of single feature description.Multifeature learning provides a feasible idea for dealing with the effects of spectral confusion.Therefore,how to characterize the HSI data and effectively alleviate the over fitting phenomenon in the training process is a key problem to be solved when designing an effective method for HSI classification with small samples size(SSS).This dissertation focuses on the problems in dimension reduction,multifeature fusion and model complexity reduction.We propose three methods to solve these problems including dimension reduction based on orthogonal constrained tensor decomposition,multifeature HSI description based on low-rank and sparse tensor decomposition,and a classifier called augmented multifeature contextual extreme learning machine.These methods improve the generalization ability of classifier and mitigate the effects of SSS.The main contribution are as follows:(1)A HSI dimension reduction method based on orthogonal constraint tensor decomposition is proposed from the perspective of dimension reduction,which reduces the complexity of the model and alleviates the over fitting problem in HSI classification.In order to obtain the spectral-spatial features during dimension reduction,tensor samples are constructed with a fixed size window to capture the contextual similarity.Aiming to improve the discriminaty of features,HSI dimension reduction is modeled as Tucker tensor decomposition with graph embedding in core tensor space.The orthogonal constraint is introduced so that the uniqueness and stability of the solution can be ensured.We proved that the optimization of objective function is equivalent to a matrix linear eigenvalue problem and an algorithm is proposed to solve it.Finally,we analyze the influence of different constraints on the performance of tensor-based HSI feature extraction.The effectiveness of this method for SSS problems is verified by experiments.(2)In order to alleviate the problem of overffitting in multifeature HSI classification,a low rank and sparse tensor decomposition method is proposed.Two problems must be considered during multifeature tensor construction: one is how to embed the contextual structure to tensors,the other is that heterogeneous features with different dimensions may result in curse of dimensionality during tensor construction.To overcome these difficulties,all features are first projected into a latent space and then stacked into tensors.The tensors are further processed by sparse and low-rank tensor decomposition in latent space.The sparse tensor can reduce the impact of outliers.The low-rank tensor can guarantee the agreements on multifeature via multigraph and preserve the structure of each distinct feature.In order to optimize the nonconvex and nonsmooth objective function,we propose an algorithm named linearized alternating direction method with adaptive penalty.Extensive experiments on three HSI benchmarks show that the OA of proposed method can reach higher than90% with only 10 training samples from each class.(3)Multifeature methods usually result in increased dimensions,which will cause the generalization ability of the classifier to decrease.To overcome this difficulty,we proposed a classifier called augmented multifeature contextual extreme learning machine,which is a semi-supervised shunt of basic contextual information.For SSS problem,staking heterogeneous multifeature into a much larger feature will aggravate the negative effects of SSS.In this context,we propose the multifeature ELM(MELM)which can make decision with multifeature in the output layer.In order to make full use of spatial context information in both the training and testing of the classifier,an augmented MCELM(AMCELM)is proposed.AMCELM not only significantly improves the classification performance on HSI but also gains better robustness to the problem of SSS and unbalanced distribution of ground objects in HSI.Experiments three HSI datasets show that the method is effective for HSI classification with SSS.