Link Day and Night: A Deep Learning Framework to Retrieve Global Nighttime Aerosol Optical Depth from VIIRS DNB

The unique Day/Night Band (DNB) sensor aboard the Visible Infrared Imaging Radiometer Suite (VIIRS) has inaugurated a new era of nighttime environmental monitoring. While most studies have primarily focused on its nighttime capabilities, largely due to the DNB mission's hallmark of enabling high-quality nighttime observation, less attention has been paid to its daytime counterparts. Due to the superior strength of sunlight, the signal-to-noise ratio (SNR) of the VIIRS DNB's daytime observations is notably higher compared to its nighttime counterparts. Concurrently, for daytime VIIRS DNB observations, a range of accompanying observations across deep blue, blue, green, red, and near-infrared bands facilitate high-accuracy AOD retrieval through advanced, optimization-based algorithms. This makes it feasible to establish a data-driven surrogate model bridging daytime Top-of-atmosphere (TOA) reflectances from the DNB with those high-quality AOD retrievals, as previous studies reveal that this panchromatic band is sensitive to the aerosol. Moreover, since the DNB operates both during the day and night, providing near-quality observations, this surrogate model could bridge also daytime and nighttime observations of DNB. Specifically, it could be applied to nighttime AOD retrieval, operating under the assumption that TOA reflectance is an inherent property of the Earth system and should display minimal dependency on the light source, whether it's the Sun or the Moon. In this study, we venture into this unexplored territory to examine the availability and the similarity between the VIIRS DNB's daytime and nighttime observations and develop a machine-learning approach, specifically a neural network, to transpose insights from daytime to nighttime. This approach expedites the retrieval of global nighttime aerosol optical depth over rural and marine areas. We validated the AOD retrieved by the neural network using CALIOP AOD for ocean surfaces and AERONET Lunar AOD for dark surfaces, focusing on the retrievals for the year 2020. The linear correlation coefficients were 0.75 for ocean retrieval and 0.72 for dark surface retrieval, demonstrating the robustness and feasibility of our neural network architecture. We will show the global nighttime retrieval greatly helps to fill the observational gaps of the atmosphere aerosol for understanding its large-scale evolution.