Research Of Method About Hyperspectral Image Classification Based On Deep Convolution Network

Hyperspectral image contains a large number of spectral bands,which is a kind of three-dimensional image data combining spectral information and spatial information at the same time.In real life,hyperspectral images can be applied to many fields.For example: in the detection of agricultural products,it is helpful for species identification;the classification of surface buildings is helpful for urban management;the recognition of pathological images can be used for disease monitoring;the classification of military maps can be applied to national defense construction.Therefore,a novel and efficient hyperspectral image classification method can play an important role in many fields.In order to improve the accuracy of hyperspectral image classification,we need to solve the difficulties of high dimension and few samples.The deep convolution neural network is helpful to mining deep-seated features in image data,so this dissertation proposes two algorithms based on deep convolution neural network to improve the classification performance.（1）From the perspective of attention mechanism,a multi-scale cross layer feature fusion network attention（MCFFN-Attention）method is proposed by using the advantages of spatial and channel attention mechanism and cross layer fusion.Firstly,the dimension of the hyperspectral image is reduced,and then the center pixel and its adjacent pixels are input into the network as a whole by using 3DCNN,and the features of different convolution layers are fused.At the same time,the low-level features are processed by spatial attention mechanism,and the high-level features are processed by channel attention mechanism.Different weights are assigned to them to optimize the feature map.The effectiveness and rationality of this method are proved by comparative experiments and ablation experiments.（2）From the perspective of enhancing receptive field,a multi-scale 3D dilated convolution Idle Block network（M3DDCIN）method is proposed by using the advantages of 3D dilated convolution Idle Block,multi-scale and channel attention mechanism.The reduced dimension hyperspectral image is input into the 3D dilated convolution Idle Block,the advantages of Idle Block and dilated convolution are used to expand the receptive field continuously,and the channel attention mechanism is used to allocate different weights to reconstruct the high-level feature map,so as to get a better feature map.Finally,experiments on three public datasets show that the classification accuracy is improved.