Hyperspectral Image Classification Based On Long Short Term Memory Network

Hyperspectral remote sensing technology is an important subject in the field of remote sensing technology,because hyperspectral data contains vast amounts of information and has unique characteristics,hyperspectral image(HSI)classification is more and more widely used in the agriculture,the geography,the meteorology and the national defense field,so it is worthy of more deep research.In the classification task,how to obtain and learn the features of data has always been the focus and difficulty,how to extract the full and effective features has a direct impact on the classification result.Inspired by the ideas of natural language processing,the focus of this paper is constructing sequence features and using deep learning framework to abstract the low level features and classify HSI to realize the integration of feature extraction and classification task.The main works of this paper are as follow:1.We proposed a manifold learning and local spatial sequential long short term memory network for HSI classification,extracting more diccriminative high-level semantic feature of HSI data.Considering that single traditional low-level feature lacking of information and different samples in local space having different effects on the classification model,we use manifold learning algorithm for dimensionality reduction to keep the local topology of image,and fuse two kinds of low-level features,then construct local spatial sequential features,finally use the long short term memory network to abstract and study local spatial sequential features and classify.This method can not only obtains more representation and discriminative high-level semantic features,but also through constructs the local space sequence to enhance positive impact of useful pixels and inhibit negative effects of useless pixels to improve the classification accuracy.2.We propose a low rank representation and non-local spatial sequential long short term memory network for HSI classification,adding global information on the basis of local spatial sequence feature and enhance the representation of high-level semantic feature.Aims at the defect that local space contains limited information,we introduce low rank representation to retain the global characteristics of the image,then find some local spaces similar to the local space of the sample in the global space of the image,and use center samples of these local spaces to construct the non-local space sequence features which not only retain the local space information but also joined the non-local space information,we can improve the classification accuracy.3.We propose a new long short term memory network using a new activation function for HSI classification.The activation function of traditional long short term memory network is Re LU function,this method proposes an improved activation function on the basis of Re LU function,and uses the low rank representation and non-local space sequence long short term memory of work 2th to classify HSI to improve the classification accuracy.