Nonreferenced Approach for Atmospheric Visibility Estimation Using Public Data and Machine Learning

Atmospheric visibility is a key metric in the safe operation of aircraft, and communicating this metric to pilots ensures operational success and compliance with federal and international law. Deployment of automated visibility estimation infrastructure is widespread near airports; unfortunately, estimation capability is sparse away from terminal airspace, mostly due to cost, hindering progress in mixed-use airspace. This paper presents advances in low-cost, low-altitude optical atmospheric visibility estimation techniques via machine learning, ideal for mixed-airspace operation of advanced air mobility vehicles. We attempt to classify images into the Federal Aviation Administration (FAA) categories of visual flight conditions, marginal visibility conditions, and instrument weather conditions via four models trained on publicly available FAA weather camera data, as well as an ensemble method. Certified visibility estimators have been shown to match expert evaluation approximately 80% of the time, which is the target accuracy of our model; the best-performing model in this study currently correctly matches the conditions predicted by these sensors 69% of the time.