Visibility Forecasting with Deep Learning

Visibility is one of the most important weather parameters that affect the operation and safety levels of transportation systems. Low visibility conditions provide highly unsafe scenarios on roads, causing accidents and jeopardizing operations. Accurate visibility forecasts play a key role in decision-making and management of transportation systems. However, the complexity and variability of weather variables makes accurate visibility forecasting a major challenge for transportation agencies nationwide. The related machine learning literature for visibility forecasting treats the problem as a classification one. However, the visibility forecasting task as a regression problem is scarce. This paper addresses the task of single step (t+1) visibility forecasting framed as a regression problem using two deep learning models: a multilayer perceptron (MLP) and a convolutional neural network (CNN). Numerical results indicate that the average root mean squared error of the models is around one mile. The CNN achieved the best average results for three hours lag and nine hours lag, while the MLP performed better for a six hours lag.