A Deep Neural Network (DNN)-Based Waterlogging Detection on Road

Abstract. Waterlogging on roads severely impacts transportation safety majorly due to inadequate drainage systems, blocked drainage channels, poor road design and construction, especially in areas with limited or no regular monitoring and maintenance, leading to increased accidents and traffic disruptions. Identifying waterlogging from the field photographic image is challenging due to poor illumination, reflective distortions, transparent surfaces, and low resolution. This study aims to identify waterlogging on rural roads using a deep learning-based semantic segmentation approach. The YOLOv11 model was trained and tested on CDAC PARAM Siddhi-AI High Performance Computing (HPC) platform. A dataset of 1000 photographic images from the PMGSY (Pradhan Mantri Gram Sadak Yojana) were sourced and annotated for this purpose. The model effectiveness was evaluated using key evaluation metrics including precision, recall, F1-score, accuracy, and Intersection-over-Union (IoU). The Deep Neural Network (DNN) model achieved a precision of 91.27%, recall of 85.95%, F1-score of 87.58%, accuracy of 96.20%, and IoU of 80.06%. The results show that training DNN model on a GPU-accelerated HPC platform significantly improves both accuracy and processing speed, which is suitable approach for waterlogging detection effectively. The output model can be utilized for deployment in national programmes such as the PMGSY National GIS, offering a rapid, cost-efficient, and scalable solution for waterlogging detection on road.