Research On Data-driven Deep Learning Models For Mixed Noise Removal Of Hyperspectral Remote Sensing Imagery

With the development of imaging spectroscopy technology,large amounts of hyperspectral datasets from different platforms(e.g.,satellite,aircrafts,and ground)have been available for multiscale earth observations and widely used in remote sensing applications,such as environment monitoring,mineral detection,and precision agriculture.However,due to the various degradation factors,including imaging environment and sensor noise,hyperspectral remote sensing images are inevitably contaminated by multiple types of mixed noises,i.e.,Gaussian,stripe,and impulse noise,which seriously degrades the image quality and usability.Hyperspectral image denoising is thus a main preprocessing step to remove the noise components and restore the real radiation of the ground objects.Traditional denoising methods,i.e.,filters and model-based optimization,rely on the precise noise models,which makes it difficult to balance the robustness and efficiency in mixed noise modeling.Recently,deep learning technology has developed rapidly,the data-driven learning manner provides a new way to overcome the dependency on expert priors,i.e.,transformation basis,image priors,and best parameters,and the difficulty of noise modeling.However,facing with the various mixed noise and complicated spatial-spectral structure of hyperspectral remote singing imagery,the lack of paired noisy-clean images and spatial-spectral integrated deep denoising network architecture is still unsolved.The quality and quantity of training dataset plays an important role in the denoising network.To alleviate these issues,this paper simultaneously takes the hyperspectral image degradation characteristics and the deep learning theory into consideration,and focus on the research on datadriven deep learning models for mixed noise removal of hyperspectral remote sensing imagery.The ultimate goal is to automatically and efficiently remove the mixed noise of hyperspectral remote sensing images,and further serve for the production and application of domestic satellites.The main research contents and innovations are summarized as follows:(1)A systematic review of the mixed noise analysis of multi-scale hyperspectral remote sensing observations and hyperspectral image denoising methods has been provided.As the deep learning technology is driven by data,this paper starts from the analysis of space-air-ground multiplatform hyperspectral images.To be more specific,the mixed noise distribution characteristics are summarized on the three dimensions,i.e.,platform,sensor,and single image,to explore the noise distributed features.(2)From the self-data-driven perspective,a Self Data-driven Denoising Network(SDDN)is proposed for hyperspectral remote sensing imagery.To alleviate the lack of paired noisy-noise images,a self-learning scheme is proposed by considering the noisy HSI itself as a training dataset,which consists of noisy bands and clean bands.To jointly restore the target noisy band and maintain the spectral consistency,a flexible multi-to-single convolutional network has been developed.More precisely,the noisy band and the past/future bands are jointly aggregating via multi-scale contextualized dilated blocks,and the separable spectral-spatial gradient convolutional unit.The framework can be applied to multi-platform observations,including satellite,manned or unmanned aircrafts,and ground.(3)From the external data-driven perspective,a Satellite-Ground Integrated Denoising Network(SGIDN)is proposed for hyperspectral remote sensing imagery.To mitigate the data dependency,a satellite-ground integrated strategy is proposed,a large number of noisy-clean pairs are generated from the ground-based hyperspectral images by randomly simulating the potential mixed noise onboard satellites.To capture the intrinsic spectral-spatial features in the HSIs,a unique CNN architecture integrated by 3D gradient convolution and multi-channel 2D convolution,residual learning is designed.Through the "Ground training and space testing" strategy,the denoising process is efficient and the single trained model is suitable for multi-source spaceborne hyperspectral sensors,including EO-1 Hyperion imaging spectrometer,the Chinese HJ-1A HSI sensor,the wide-range hyperspectral imager onboard the Chinese SPARK micro-nano satellite,and the Zhuhai-1 hyperspectral imager.