Feature Extraction And Classification For Hyperspectral Image Based On Convolution Neural Network

Feature extraction and classification of data have been the hot spot in research of remote sensing technology. Using linear or nonlinear equation artificially designed or specified extracted features, the existing feature extraction method is mainly aimed at the characteristics in a certain respect of the target object. This kind of artificial feature selection process often need professional knowledge and experience, and spend a lot of time. However, the extracted features can not fully express the complex internal structure and empty spectrum information of hyperspectral data. For deep learning, it can automatically learn the features that the task need by computers, and implement the steps in the process of model training, which help improve the classification accuracy.This thesis starts with the characteristic of hyperspectral data, combining the convolutional neural network model based on deep learning, and using multiple convolution and pooling layers to extract the nonlinear features from hyperspectral data which are also highly invariant to the translation, scaling and rotation for improving the precision. The main research work in this paper focuses on the following aspects:Firstly, according to the characteristics of image with spectrum, we explore the applicability of the deep convolution network for the hyperspectral data feature extraction and classification. When acquiring the space information of hyperspectral data, we can get a continuous spectrum curve on a pixel. It makes the hyperspectral data with high dimension and large amount of data, and deep learning model is suitable for the characteristics of this data. So this thesis use spectral information, spatial information and spectral information of hyperspectral data, and construct respectively deep convolution convolution kernels neural networks based on the one-dimensional, two-dimensional and three-dimensional, to express features hierarchically and formally. Classifying the hierarchical features is superior to other methods’ results of feature extraction and classification.Secondly, in view of the imbalance problem of high dimension with limited training samples, this paper respectively introduce L2 regularization in one-dimensional convolution model to modify the original cost function, and add Dropout in 2D and 3D model layer to sparse the activation of each layer of the network units, to avoid over fitting phenomenon in the process of modeling. And using the unsaturated nonlinear function Re LU instead of the original Sigmoid activation function, it can greatly improve the convergence speed of the model and reduce the complexity of the model.Finally, with the less training samples, the model of one-dimensional deep convolution based on spectral information can not provide stable classification results. In order to further improve the classification performance of the model, an ensemble model is proposed in this paper based on the depth of the random feature selection. Data from two groups of the experimental results show that compared with other classification accuracy of the method, the method is a more competitive solution. The experimental results of the two datasets show that the method is a more competitive solution compareing with other classification accuracy of the method.