Efficient Deep Learning Methods for Automated Visibility Estimation at Airports

The goal of this paper is to design and develop new visibility estimation methods for airports based on deep neural networks. Their input consists of partially annotated image meteorological data, especially camera images from automatic meteorological stations. These have been processed, normalized, cleaned, and annotated according to meteorological observation codes and geographic data, with manually entered values. We have experimented with different methods of image processing, leveraging deep learning in order to find and finetune a lean, yet effective neural network architecture capable of providing the desired visibility estimation quality.