Day time and Nighttime Dust Event Segmentation using Deep Convolutional Neural Networks

Airborne dust is known to have detrimental effects on human health, the environment, and aviation. Earth-observing satellites have been used to monitor dust events using visible and infrared bands, but the presence of clouds and smoke makes it a difficult phenomenon to identify. Moreover, nighttime dust has similar radioactive properties to that of cooler underlying surfaces. We propose a dust detection algorithm that uses false color EUMETSAT Dust Red-Green-Blue (RGB) imagery (dust RGB). The false-color imagery aims to enhance the dust detection process in both daytime and nighttime by using band differences of GOES 16 satellite. A deep learning-based segmentation model is trained to detect dust using the same bands used to create the dust RGB imagery. The model has high accuracy in detecting dust events and does not require large amounts of training data. The model also performs well in both day time and night time situations.