Coloring panchromatic nighttime satellite images: Elastic maps vs. kernel smoothing and multivariate regression approach.

Artificial light-at-night (ALAN), emitted from the ground and visible from space, marks human presence on Earth. Since the launch of the Suomi National Polar Partnership satellite with the Visible Infrared Imaging Radiometer Suite Day/Night Band (VIIRS/DNB) onboard in late 2011, global nighttime satellite images have considerably improved in terms of spatial resolution, quantization, saturation, and light detection limits. However, VIIRS/DNB images remain panchromatic, reporting aggregated light emissions in the 500-900nm range. Although multispectral satellite images are also available, such images at present are either commercial or free, but sporadic. In this paper, we use different machine learning techniques, such as linear and kernel regressions and elastic map approach, to transform panchromatic VIIRS/DBN images into RGB. To validate the proposed approach, we analyze nighttime satellite images available for eight urban areas worldwide. The analysis links RGB values, obtained from International Space Station (ISS) photographs, to panchromatic ALAN intensities, obtained from VIIRS/DBN images and combined with pixel-wise proxies for land use types. During the analysis, each dataset is used for model training, while the rest of the datasets are used for model validation. We compare the models' performance using several performance measures for training and testing sets. As the analysis shows, the modelled RGB images demonstrate a high degree of correspondence with the original RGB data. Yet, estimates, based on linear and non-linear kernel regressions, appear to provide better correlations and lower WMSEs, while RGB images, generated using the elastic map approach, appear to provide better contrast similarity and better consistency of predictions. The proposed method demonstrates its utility and can thus be used for obtaining seamless RGB coverages using panchromatic VIIRS/DBN data.