Hyperspectral Image Classification Based On Convolutional Neural Network By Spatial-spectral Feature Fusion

Hyperspectral image classification technology is a significant part of the field of remote sensing.It is popularly utilized in agriculture,biomedicine,military and other fields,and has important value.However,the extremely high spectral resolution of hyperspectral data brings the possibility of hyperspectral data classification,but also brings problems such as large data volume and spectral information redundancy.In addition,the spatial structure of hyperspectral is complex.These problems have brought huge challenges of hyperspectral classification.Convolutional neural network CNN recently has performed outstandingly in the field of hyperspectral image classification and has been popularly used.At present,the spatial and spectral features extracted by deep learning in the hyperspectral image classification method are not very separable,and requires a large number of labeled samples,but the actual number of training samples is limited.These problems affect the classification accuracy.In view of these issues,this article conducts the following research.1.This topic has studied in depth the classic hyperspectral image classification methods.On this basis,it analyzes the shortcomings of shallow learning methods,and then introduces deep learning theory,and introduces deep learning models and training methods in detail.The model of convolutional neural network is introduced in depth base on deep learning theory research to lay the foundation for the subsequent construction of a hyperspectral image classification model based on convolutional neural networks.2.Aiming at the problem of insufficient spatial and spectrum features extracted by traditional CNN models and the disappearance of gradients in the construction of deep convolutional networks,a classification method based on spatial and spectral feature enhancement is proposed.The method uses residual learning to design spatial and spectral feature enhancement modules,which not only strengthen the extraction of spatial and spectral information,but also deepen the depth of the network model to more fully mine high-level semantic features.In addition,the method fuses the extracted spatial and spectral features with the original information,and the fused features are input into a shallow three-dimensional CNN network to achieve fusion extraction of different levels of features,enhancing the separability of high-level and abstract semantic features.Experimental results show that this method uses fewer parameters to construct a deep network,can fully extract spatial and spectral features,enhance the separability of features,and improve the classification accuracy when there are relatively few samples.3.Aiming at the problem that the classification accuracy of hyperspectral images based on CNN is low in small samples,this paper proposes a classification method combining active learning and three-dimensional CNN.First,the method uses a network trained with a small number of known samples to obtain the value of unknown samples,and the active learning algorithm selects samples that greatly improve the performance of the classifier through value comparison.Then,an automatic labeling algorithm is used to label the selected unknown samples based on the principle of local spatial consistency.The algorithm uses the convolutional neural network and the neighbor propagation algorithm to determine the label of the sample.Finally,the above labeled samples are added to the training set to improve the classification performance of the network.The experimental results show that the method selects samples that greatly improve the classifier through active learning,and the automatically labeled samples are correct,reducing the number of training samples.This method raises the classification accuracy under small samples.