Dehazing in hyperspectral images: the GRANHHADA database.

In this study, we present an analysis of dehazing techniques for hyperspectral images in outdoor scenes. The aim of our research is to compare different dehazing approaches for hyperspectral images and introduce a new hyperspectral image database called GRANHHADA (GRANada Hyperspectral HAzy Database) containing 35 scenes with various haze conditions. We conducted three experiments to assess dehazing strategies, using the Multi-Scale Convolutional Neural Network (MS-CNN) algorithm. In the first experiment, we searched for optimal triplets of spectral bands to use as input for dehazing algorithms. The results revealed that certain bands in the near-infrared range showed promise for dehazing. The second experiment involved sRGB dehazing, where we generated sRGB images from hyperspectral data and applied dehazing techniques. While this approach showed improvements in some cases, it did not consistently outperform the spectral band-based approach. In the third experiment, we proposed a novel method that involved dehazing each spectral band individually and then generating an sRGB image. This approach yielded promising results, particularly for images with a high level of atmospheric dust particles. We evaluated the quality of dehazed images using a combination of image quality metrics including reference and non-reference quality scores. Using a reduced set of bands instead of the full spectral image capture can contribute to lower processing time and yields better quality results than sRGB dehazing. If the full spectral data are available, then band-per-band dehazing is a better option than sRGB dehazing. Our findings provide insights into the effectiveness of different dehazing strategies for hyperspectral images, with implications for various applications in remote sensing and image processing.