A Hyperspectral Super-resolution Method Based On Spatial Spectral Depth Feature Fusion

Hyperspectral remote sensing images have the characteristics of high spectral resolution.However,due to the limitation of imaging sensors,hyperspectral images have lower spatial resolution.It greatly affects the application of hyperspectral images,so improving the spatial resolution of hyperspectral images is a hot research topic.Hyperspectral remote sensing image super-resolution technology uses software methods to reconstruct high-resolution images using low-resolution hyperspectral images.Traditional matrix decomposition methods rely on highresolution RGB images to provide additional spatial information,otherwise the reconstruction accuracy is insufficient.The deep learning method learns features directly from the data.It can describe the deep features of the data,and can achieve high reconstruction accuracy through these internal information.In this paper,a CNN is constructed to solve the HSI SR problem,and proposes an optimization method that iteratively optimizes the reconstruction results based on the super-resolution reconstruction of the CNN.Main tasks as follows:(1)In this paper,a cascaded network architecture(SS-UNet)is proposed for SR of HSIs.It introduces 3D convolutional neural network to improve the super-resolution reconstruction effect of hyperspectral images.The network calculation process is divided into three steps: First,the HSI features pass through the contracting path,the spectral features are continuously reduced,and the spatial features are continuously refined.Second,use more 2D convolutions to further extract the depth space information of the shrunk features to further refine the spatial feature.Finally,the feature is continuously reconstructed through the expansive path to improve the spectral dimension.The experimental results on the HSI data set show that SS-UNet can improving the super-resolution reconstruction accuracy of HSI.It has better performance,can reduce the amount of calculation and memory usage.(2)Based on the alternating multiplier method and CNN,this paper proposes an optimization algorithm for HSI SR.Deep learning methods can learn the internal laws and deep representations of the data.But the CNN model learned through the training process is uninterpretable.It is difficult to optimize the parameters of the trained model.The objective optimization function we constructed includes three parts: data guarantee term,model guarantee term and gradient specification term,and uses alternating multiplier method to optimize the reconstruction result.The design of the objective optimization function fully considers the impact of input data and model on the SR reconstruction of HSIs.It makes the optimization result not only have data consistency with the low-resolution input image,but also maintain model consistency with the trained model.The experimental results prove that the optimization method proposed in this paper can improve the reconstruction accuracy based on the model output.(3)On the basis of the SS-UNet in this article,this paper develops a hyperspectral image super-resolution system based on the QT graphical user interface.The basic functions of data selection,preprocessing,super-resolution reconstruction,and quantitative evaluation of reconstruction effects are realized.This system has good interactivity and cross-platform.