A Denoising Method for Hyperspectral Remote Sensing Image Based on Unsupervised Adaptive Learning

Abstract

The invention relates to a hyperspectral remote sensing image denoising method based on unsupervised self-adaptive learning. Aiming at the problem that the quality degradation difference between the simulated image and the real image reduces the generalization of the model, the invention proposes an unsupervised adaptive learning strategy, performs pre-training on high-quality ground images, and designs a discriminator to model noise. When processing real images, the discriminator fine-tunes the denoising parameters to improve the generalization of the model on real images. In the present invention, a deep denoising network based on spatial-spectral residual and a discriminator based on global information modeling are designed in the model to fully mine the hyperspectral depth prior. The invention can solve the problem of simulating training data in deep learning denoising of hyperspectral remote sensing images, reduce the dependence of the deep learning model on the simulated training data, and effectively improve the applicability and accuracy of hyperspectral denoising.

Description

technical field [0001] The invention is based on the field of remote sensing image technology processing, and in particular relates to a hyperspectral remote sensing image denoising method based on unsupervised adaptive learning. Background technique [0002] With the increase of the spectral resolution of sensors, hyperspectral remote sensing has been developed rapidly. Hyperspectral satellites can quickly acquire images of hundreds of bands in the same area, and have the advantage of integrating maps. However, in the process of data acquisition by hyperspectral remote sensing satellites, due to harsh atmospheric conditions, imperfect calibration process and the defects of the sensor itself, hyperspectral remote sensing images will inevitably be polluted by various noises, resulting in stripe noise, Degradation problems such as random noise and reduced image contrast severely limit the availability of image data. In the image data with mixed noise, the image information i...

AI Technical Summary

Problems solved by technology

[0003] At present, the existing hyperspectral mixed noise removal methods have great limitations in practical applications

Among them, the filtering-based method is limited by the fixed transformation domain, it is difficult to fully mine the geometric characteristics of hyperspectral data, and it is difficult to determine the appropriate cut-off frequency in practical applications; the method based on the regularization model is based on the noise characteristics and potential noise-free The artificial assumption of image features makes it difficult to accurately model the complex noise in massive hyperspectral data, and when more constraints are added, the solution process is very complicated, the calculation efficiency is low, and it is difficult to meet the requirements of practical applications; based on deep learning The method has high computing efficiency, but its training process relies on a large number of noise-noise-free image pairs. Since airborne and spaceborne hyperspectral remote sensing platforms usually cannot obtain real image pairs, deep learning models are mostly carried out on simulated image pairs. Training, which reduces the generalization of the model on real data

Images