Research On Hyperspectral Images Super-Resolution Reconstruction Based On Deep Learning

High-spectral images have a wide spectral field to accurately detect fine details of the landform.It is extensively used in areas such as ecological research,geological exploration,and remote sensing monitoring.High-spectral imaging sensors often require a larger spatial range to collect photon data in order to ensure the SNR on the narrow band,however,this reduces the spatial resolution of the hyperspectral image.Due to hardware constraints,high-spectral image superresolution reconstruction is an important direction in remote sensing image processing,which cannot be solved in a short time.In this research,we use deep learning to combine high-spectral images with multi-spectral images of high spatial resolution to realize the task of super-resolution reconstruction and do the following work.(1)In view of the problem that existing methods cannot fully extract spatial features and spectral features of remote sensing images,a hybrid attention module integrating spatial attention and spectral attention is designed in this paper.The spatial attention module can enhance the representation ability of spatial features and avoid the loss of local regional details.The spectral attention module screens the numerous bands of remote sensing images to capture valuable spectral information while ignoring redundant useless information.Therefore,this paper proposes a hyperspectral image super-resolution algorithm based on a hybrid attention mechanism,which not only helps to reconstruct spatial features but also emphasizes the correlation between spectral bands,reduces the redundancy and overlapping between bands,so that the final feature information obtained by the network is more obvious and complete.(2)Due to the imaging differences of different sensors,the scale difference of the multispectral and hyperspectral images obtained from the same object is large,and the scale usually needs to be unified before image fusion,which may lead to information loss or accuracy reduction,affecting the relationship between bands in network learning.Moreover,most algorithms only extract image features on a single scale,making it difficult to retain the original image features.To address the above problems,this paper proposes a hyperspectral image super-resolution algorithm based on a multi-scale convolutional neural network.The convolution layer branches designed in the network structure are equipped with various convolution kernels to extend the receptive field of feature extraction,and the network structure gradually reduces the feature size of multispectral images at different scales and integrates the features with hyperspectral images,thus establishing a mapping model from low-resolution hyperspectral and multispectral images to high-resolution hyperspectral images.In this research,two different well-known high-spectral public datasets were used to compare with related algorithms in recent years,and erasure experiments,analysis experiments of objective indicators,and subjective visual evaluation were conducted to verify the practicality of the algorithm.