Automatic Fog Detection and Visibility Estimation From Camera Images Using Deep Learning Features For Aviation Operations Involving Uncrewed Aircraft Systems (UAS)

The degradation of visibility caused by foggy weather conditions can cause major issues that pose an ever-present challenge to aviation operations, especially small Uncrewed Aircraft System (UAS) since they are more vulnerable to weather conditions. Therefore, accurate prediction of reduced visibility conditions is crucial to ensure efficient and safe UAS mission planning. Obtaining forecast initialization and validation data is difficult due to the small scale of many fog events. Additionally, remote sensing observations are difficult due to the low altitude of fog. Traditional visibility measurement techniques, such as visibility sensors and meteorological observers using landmarks, are insufficient to provide the required fog information over a large area since observations are limited due to sensor and maintenance costs, instrument issues, required maintenance, and availability of the meteorological observers.

A method for determining the existence of fog and automatically deriving visibility is described that uses digital cameras as potential sensors. The cameras are a reliable alternative for identifying areas with low visibility conditions when the processing of image data is used. The approach is to use a deep neural network to determine the presence of fog and make a real-time estimation of the visibility. For visibility estimation, a labeled data set from camera images is developed for various weather conditions over North Dakota, especially during fog events. A set of features are extracted from the camera images, such as the number of edges, brightness, and the transmission of the image dark channel, to develop a fog detection algorithm and a regression model to estimate the visibility. The regression model, in addition to the image features, are utilized for training the deep neural network. The evaluation data set is used to test the network and validate its performance. The performance of the neural network will be evaluated against operational models used in Aviation.