Hyperspectral Remote Sensing Image Classification Based On 3D Deep Learning Network

With the continuous advancement of remote sensing technology and newer development of sensors,the research value of hyperspectral remote sensing images has been increasing.Hyperspectral image classification,as one of the most popular research fields,plays an important role in several different industries.Compared with traditional multispectral images,hyperspectral image maps contain richer wavelength and spatial information of features and exhibit stronger potential for detecting target features.In recent years,deep learning has provided a new and effective way for hyperspectral image classification due to its ability to achieve end-to-end and adaptive feature extraction,and has shown stronger feature extraction capability in practical applications.In this thesis,by comparing and analyzing the effects of classical deep learning models in hyperspectral image classification,we study and improve the data band selection and deep learning model optimization,and propose a hyperspectral image classification method based on 3D deep learning network,which finally improves the effect of feature classification.In terms of band selection,the standard score evaluation index of the bands of hyperspectral data is proposed to filter the effective bands and eliminate the bands with poor quality in view of the impact of redundant bands on data quality.In terms of model optimization,three improvements are made:（1）a 3D convolutional network model is proposed to use the spatial information of hyperspectral data for feature extraction;（2）the computational effort is reduced by using convolutional operations instead of pooling operations to reduce the deep learning network parameters;（3）the number of parameters of the network model is greatly reduced by optimizing the residual network model.This study shows significant improvement in both accuracy and efficiency of hyperspectral remote sensing image classification.Tested by hyperspectral images Pavia Scene and Pavia University datasets,the convolutional neural network in this thesis achieves 99.01% and 95.99% accuracy,respectively.In the small sample training test sets Pavia University and Salinas,the pyramidal residual network proposed in this study achieves 96.86% and 96.49% classification accuracies,respectively.The improved convolutional neural network reduced the parameters by about 40% and the training time by about 74%,and the model training speed was improved while the final classification accuracy could be maintained at a more desirable level.