Research On Classification Technology Of Hyperspectral Remote Sensing Image Based On Deep Learning

With the continuous development of remote sensing technology and hyperspectral measurement methods,hyperspectral remote sensing image classification methods have become one of the hot issues in research.The traditional feature extraction and classifier framework,which uses artificial extraction features,is often not comprehensive enough to fully utilize the spatial-spectral information of hyperspectral images.Combining with the widely popular deep learning technology methods,this paper starts from the characteristics of hyperspectral image data,and uses 3D convolution to fully extract the spatial-spectral features.Adaptive extraction of features without the need to manually design features.It implements an end-to-end model.Compared with the existing models for processing hyperspectral images,there are many improvements in accuracy and speed.The main research contents and achievements include the following four aspects:(1)A 3D-CNN based spatial-spectral information combined hyperspectral image classification method is implemented.Based on the 3D-CNN structure,using the characteristics of the hyperspectral remote sensing image data cube,the model achieves the full fusion of spectral and spatial information,and extracts the discriminative features that are important in the classification.The network structure considers the lack of label data of hyperspectral remote sensing images,and also considers the balance between high-dimensional characteristics of hyperspectral imagery and model depth.It makes full use of the semantic information provided by the combination of the spatial and spectrum.(2)A 3D-DenseNet hyperspectral image classification method based on 3D-CNN and dense connection structure is implemented.3D-DenseNet enhances feature transmission through dense connection structure,encourages feature reuse,and provides information flow in the network.Combining the excellent characteristics of 3D-CNN and dense connections,deep extraction of spatial-spectral features is achieved.(3)The 3D-SE-DenseNet hyperspectral image classification method that introduces the attention mechanism is realized.In view of the sparse characteristics of hyperspectral data,the sample distribution is uneven in space,and there is a large amount of high-dimensional redundant information in the spectral dimension.Therefore,the Squeeze-and-Excitation module(SE module)is introduced.It correlates the convolutional feature maps between different channels,accelerates the understanding of the ground type by the depth model,effectively controls the total amount of parameters,and improves the classification accuracy,especially when dealing with small sample data.(4)A hyperspectral image classification method based on feature aggregation method for deep feature aggregation networks is implemented.From the perspective of aggregation view,combined with residual structure and dense connection structure and feature fusion mechanism,the network can be transmitted deeper.It is capable of extracting depth features.When faced with the characteristics of different hyperspectral data,the corresponding aggregation method can be selected more flexibly.