Applying Artificial Intelligence for the Detection and Analysis of Weather Phenomena in Vehicle Sensor Data

Abstract

Weather predictions arise from observatory stations on fixed locations, forming a nationwide grid. The low resolution of this grid does not allow for the prediction and discovery of local road weather conditions. This paper aims to identify weather conditions on a high-resolution scale by applying machine learning on vehicle sensor data. The model classifies anomalous samples in time series data into a road weather condition. We examine how Decision Trees can be applied to classify anomalous vehicle behavior into weather phenomena. It also specifies which preparation steps on sensor observations are advisable before a model is applied. We constructed numerous Random Forest and Gradient Boosted Tree classifiers to classify anomalies of real-world vehicle data. The grid search performed on classifier hyperparameters and input configurations shows that a well-considered feature selection and filtering has a significant impact on the accuracy.