Research On Hyperspectral Image Classification Based On Deep Correlation Feature

Hyperspectral image(HSI)is a kind of composite information image data,which contains both spatial information and spectral information.Compared with natural images,hyperspectral images have richer spectral features.Therefore,Hyperspectral remote sensing(HRS)has great application value and is widely used in geological exploration,environmental monitoring,military surveillance and agriculture.In recent years,with the development of optical sensors,the imaging quality of HSIs is getting higher and higher.Reasonable analysis of HSI data can accurately distinguish the classes of ground cover,therefore,HSI classification has become one of the most important applications in HRS.At present,deep learning has made great breakthroughs in the field of HSI classification with its strong feature representation ability,but there are still some shortcomings,including:(1)it has a strong dependence on the amount of training data,and it does not perform well in the case of small samples;(2)it ignores the correlation characteristics between samples;(3)it lacks guidance for the design and construction of the network.Aiming at the above problems,this thesis sorts out the HSI classification algorithm suitable for small sample size and verifies the effectiveness of sample correlation characteristics for improving classification accuracy,and proposes two HSIs classification algorithms based on depth correlation features.In order to achieve satisfactory classification results in the case of small samples,this thesis uses a label diffusion training method,so that only a very limited number of training samples can be used to fully train the model.Besides,compared with the mainstream training method based on local patch blocks,the label diffusion training method used by the text avoids a large number of double calculations in the forward propagation process,and greatly improves the inference speed of the model.Considering the integrity of spectral images,there is a wide range of connections between pixels.To capture the correlation features between pixels,this thesis proposes a novel HSI classification algorithm that combines deep learning and sparse representation.This algorithm simulates the solution of sparse representations through a deep learning network and solves the abundance of sparse representation in the forward propagation process of deep learning.The solved deep abundance feature can effectively represent the correlation features between samples,and the deep abundance The features are fused with the original data,and input into the feature extraction module with U-Net as the back-bone to output the final probabilistic classification map.The desirable results on three widely used datasets validate the effectiveness of the proposed method.In order to capture more robust and discriminative correlation features between samples,this thesis further proposes a HSI classification algorithm based on composite kernel network.This method introduces the generalization kernel method,and calculates the generalization composite kernel feature matrix in the forward propagation process of deep learning.Each sample feature in the extracted generalization kernel feature contains the composite kernel function calculation between this sample and all other training samples.Combined kernel computations are performed to be more robust and discriminative.Similarly,a U-Net skeleton feature extraction module is used to obtain the final probabilistic classification map.Through experimental verification,this algorithm can fully exploit the potential correlation features between samples and achieve satisfactory accuracy.