Research On The Noise Reduction Algorithm Of Hyperspectral Images Spectral Domain Based On Spectra Correlation

With the increasing resolution of detectors, hyperspectral remote sensing has been widely used in the field of economy, military, environmental protection etc. But the hyperspectral images(HSIs) will be degraded by the noise inevitably when it is acquired or transmitted. The noise exists in both spatial domain and spectral domain, so the noise reduction(NR) must do at both domains.Utility the correlation of bands and the spectra of nearly pixels can get a better NR result. When do the NR in spatial domain, the correlation of bands have been paid attention, but in spectral domain, the correlation of spectra has been ignored.In this paper, a novel NR algorithm is provided which take full advantage of the correlation of spectra when do NR at spectral domain. The algorithm utility the principal component analysis(PCA) to remove the correlation of spectra, and to avoid the biased estimation of PCA, use homogeneous region segmentation to divide the HSIs, which give a more accurately regions with high spectral correlation, meanwhile, the merge of undersize regions, have made to that to avoid the influence of noise and make the algorithm have more robustness when the parameter is small. According to that, a whole wavelet based NR algorithm is given, and NR experiments have made on simulated and real HSIs. The results have been compared with that of four NR algorithms. To simulated HSIs, the evaluation of NR has been made from the total signal to noise ratio, visual effect and MSSIM, MRMSE and MNCC. The proposed algorithm get the best results at any evaluation criterion. To the real HSI, the evaluation of NR has been made from the classification accuracy. The overall accuracy reached 92.46%, which is the best of five NR algorithms and enhanced nearly 8% compared to the noisy data; the Kappa reached 0.9141.Finally, give out the analysis of the parameters influence of the algorithm. To the wavelet basis and threshold selection function, the proposed algorithm can get a satisfied NR result at any choices, the most difference is just 0.1db. To the segmentation parameter, the modified homogeneous region segmentation algorithm can get a 0.8db improvement when the parameter is small compared with the original algorithm. Simply apply PCA to the spectra can also get a better result compared to the other four algorithms.