Integrating convolutional neural networks and explainable AI for enhanced winter road surface conditions classification using stationary RWIS imagery

Adverse winter weather significantly compromises driving safety and mobility in regions such as Canada and the northern United States. This study addresses these challenges by utilizing stationary Road Weather Information Systems (RWIS) equipped with cameras that capture complex scenes, making road surface condition (RSC) classification particularly challenging. Unlike previous studies that required manual cropping main road pavement, we applied convolutional neural networks (CNNs) directly to full stationary RWIS imagery to validate their effectiveness and generalizability for real-world winter road maintenance (WRM) applications. Our study focused on four key aspects: (1) validating CNN on stationary RWIS images; (2) analyzing the impact of camera angles using explainable artificial intelligence (XAI) techniques to gain insights into the CNN's decision-making processes; (3) examining the effects of image resolution on performance to identify an optimal balance between accuracy and computational efficiency; and (4) exploring how varying camera numbers and angles contribute to model’s performance. The developed CNN demonstrated strong performanc, with accuracy, balanced accuracy, and F1 macro score consistently exceeding 96%. Our findings indicate that optimizing camera orientation substantially enhanced the model's focus on relevant features that align with human interpretation. Reducing background complexity and increasing road captures from different perspectives further enhanced model focus. Furthermore, increasing image resolution up to 224×224 improved performance, although gains beyond this point were marginal while computational costs rose substantially. This comprehensive evaluation demonstrates the high potential of using stationary RWIS imagery for RSC classification with CNNs, ultimately leading to improved WRM operations and traffic safety during winter.