Hyperspectral Images Classification Based On Non-overlapping Sampling

The most important feature of hyperspectral images is that they contain both spatial features and spectral information,which enables them to identify ground objects more accurately.Because the introduction of spatial information can greatly improve the classification accuracy of hyperspectral images,the spectral-spatial classification method has attracted the attention of a large number of researchers,but the random sampling method will lead to the high overlap of training samples and test samples in the spatial domain and feature domain,resulting in the false classification accuracy.Recently,a new non-overlapping sampling method is proposed which is more suitable for the practical application environment.To solve this problem,this paper conducts an in-depth study on the classification methods under the non-overlapping sampling strategy,and proposes two methods based on improved collaborative representation and a deep learning model to improve the classification accuracy of hyperspectral images under the non-overlapping sampling.The specific contents are as follows.1.A classification method based on collaborative representation and set to set distance is proposed.Firstly,the training sets and the test sets are obtained by non-overlapping sampling and superpixel segmentation respectively.Then,the training set is compressed into a more compact matrix to reduce the computational complexity,and the kernel trick is introduced to make samples approximately linearly separable.Finally,each test set is modeled as a convex hull and represented by all training sets collaboratively,and then the classification is completed by calculating the distance between the test set and each training set.2.A collaborative representation method based on spatial regularization constraint is proposed.The model effectively fuses spectral and spatial information.In order to fully mine spatial characteristics,two regularization terms are designed in the classification model based on collaborative representation: one is the constraint on spectral similarity,the other is the constraint on spatial similarity.Since the improved objective function is still an unconstrained optimization problem,we can simply use the least square method to find an analytic solution to the representation coefficient.Finally,the label is determined by minimizing the prediction error of the test sample.3.A classification method based on prototypical network is proposed.Firstly,a metric space is learned by training a two-dimensional convolutional neural network,and the metric space is parameterized by the two-dimensional convolutional neural network.Then,the model is updated by calculating the softmax function of the Euclidean distance between the training sample and each prototype of class.Finally,after training the model,the saved model is directly used to extract the characteristics of the test set and complete the classification.In this paper,the public hyperspectral data set is used for experimental verification,and the experimental results fully prove the effectiveness of the three hyperspectral image classification methods under non-overlapping sampling.