Hyperspectral Image Classification With Cascaded Support Vector Machines And Gabor Convolutional Networks

Hyperspectral imagery contain dozens to hundreds of bands from visible to infrared spectrum.In the last 10 years,hyperspectral image classification has become an active research subject in the hyperspectral field.The main purpose of hyperspectral image classification is to accurately and efficiently classify features of interest in a hyperspectral image.Hyperspectral image classification have been successfully applied to many fields: ecological science,geological science,hydrological science,precision agriculture,military applications,etc.This thesis is mainly for the research of hyperspectral image classification algorithms.By combining some popular classification algorithms with excellent performance,we have developed some excellent algorithms for hyperspectral image classification based on its inspiration and basis.First,by using a new deep neural network algorithm idea based on cascade structure,an algorithm suitable for hyperspectral image classification is proposed.Though deep models have obtained impressive results for HSI classification,lots of labeled samples are often needed to tune the abundant parameters of deep models.Zhihua Zhou zhihua proposed a new deep neural network method(gc Forest).Compared with deep neural network,gc Forest is much easier to train.This method adopts cascade structure,and obtains highly competitive performance in a wide range of tasks.Based on this model,we propose a cascaded support vector machine structure.In this structure,we use support vector machine instead of random forest.This is because the multigrained scanning is not necessary for hyperspectral image classification processing.The dense connected structure is also introduced to increase the information flow between layers.In addition,after forming the enhanced feature,two round major voting based on multi-scale superpixel segmentation is used to introduce spatial information.Next,a new classification model is proposed based on the cascaded support vector machine,which combines multiple feature extraction methods with the cascaded support vector machine.This model adds different feature extraction methods to each layer of the cascaede support vector machine for feature extraction of input.Different feature extraction methods are complementary to each other,so the information extraction of hyperspectral images is very sufficient,and the classification accuracy of hyperspectral image is improved.Finally,a new network structure is applied to hyperspectral image classification.This structure introduces gabor filter into the convolutional neural network,and the convolutional gabor orientation filter is formed by using Gabor filter to adjust the convolution filter.By replacing the convolutional layer in the convolutional neural network with the convolutional gabor orientation filter,the gabor convolutional network we use is obtained.The network can enhance the adaptability of deep learning features to orientation and scale,and experiments have proved that the network has excellent performance in hyperspectral image classification.