Improving Generalizability of Road Condition Classification Models for Department of Transportation Camera Images

The New York State Department of Transportation (NYSDOT) has a large network of publicly available roadside cameras (511ny.org) that are used by both the public and NYSDOT to monitor road conditions, especially during winter storms. In an effort to automate the manual task of monitoring cameras, convolutional neural networks are trained to classify images into one of six road surface categories: severe snow, snow, wet, dry, poor visibility, or obstructed. These road-surface models are successful at identifying the road conditions on unseen images (the validation dataset) when all cameras and dates are represented in model development (the training dataset), achieving validation accuracy of ~90%. However, validation accuracy drops to ~60% when evaluated on new dates or cameras, indicating the need to explore additional methods to achieve model generalizability. Improving model generalizability is crucial in developing a tool that can be used in future winters to inform resource allocation and driving safety across the state of New York on thousands of new cameras in different and changing environments.

The model’s ability to generalize to unseen data is impacted by each decision made in the model development process, such as: sampling and labeling images, selecting cameras to include in training, including pre-processing techniques, and choosing a model architecture and its hyperparameters. A set of experiments were run to understand how each of these decisions impacts model performance from the perspective of improving model generalizability. This presentation will share the design and results from these experiments with an emphasis placed on model performance on new cameras, and it will include an outlook on future work needed for eventual model use across New York State.