| In the 21 st century,hyperspectral remote sensing technology has been widely used in agriculture,environment,medical,biological and other fields,especially in crop monitoring and resource environment monitoring.Based on the abundant spatial and spectral information of hyperspectral data,crop growth and environmental pollution can be accurately obtained.However,due to the influence of the external environment and the limitation of the instrument itself,it is not so simple to obtain high-quality hyperspectral images.First,the existence of noise seriously affects the subsequent application of hyperspectral data,such as classification,disambiguation,inversion and so on.Therefore,hyperspectral denoising is an indispensable step.Second,hyperspectral acquisition requires harsh weather conditions and expensive instruments,which greatly increases the acquisition time and acquisition cost of hyperspectral data.Therefore,it is extremely necessary and meaningful to recover hyperspectral information from simple natural images.Based on the above two points,this paper uses some classical machine learning algorithms and the deep learning technology to study hyperspectral image restoration algorithms.The main contributions are as follows:1.Based on the properties of spatial correlation and spectral correlation of hyperspectral images,Mahalanobis distance is used to establish the similarity weights of non-local spatial spectral domain,which are used to represent the statistical distribution of pixels in non-local region of hyperspectral images.This distribution can be integrated into the PCA algorithm to achieve denoising by selecting the principal component.2.Although the amount of hyperspectral image data is relatively large and the update period is relatively fast,the high-spectral data can be quickly denoised in batches by using the Generative Adversarial Networks(GAN).The adversarial learning of Generative model and Discriminative model can not only effectively suppress the noise in the hyperspectral image,but also preserve the detailed texture information of the image.3.Using the powerful image generating ability of GAN,a SNGAN network is designed to recover the spectral information of hyperspectral images from RGB natural images.To some extent,it solves the problem that the discriminator of GAN network is not easy to converge.In this way,not only the hyperspectral data can be easily obtained,but also the high-quality noise-free hyperspectral images with the same spatial resolution as RGB images can be quickly obtained,which greatly improves the application scope of hyperspectral images.The above experimental results show that the non-local spatial spectral PCA algorithm not only suppresses a lot of noise,but also preserves the boundary information in the image.At the same time,the article finds that the deep learning algorithm can be applied to remote sensing,especially complex hyperspectral image restoration,which can not only save manpower and material resources while obtaining high-quality hyperspectral data,but also make hyperspectral images suitable for the background of the era of big data,and promote hyperspectral images to truly embark on a commercial road. |
