Intelligent Flood Detection using Traffic Surveillance Images based on Convolutional Neural Network and Image Parsing

Abstract

An intelligent flood detection system is developed from an existing traffic surveillance structure. Images are captured from closed-circuit television (CCTV) with actual setting conditions - (a) normal, raining and flooding, and (b) day and night. The proposed system applied scene parsing method to avoid the impact of varying the physical setting of CCTV structures. This image parsing method uses pre-trained model, DeepLabv3, to detect objects common to traffic CCTV images such as road and vehicles. Supervised learning is performed to detect floods based on a convolutional neural network (CNN) model. The CNN model is validated ten times by training and testing it with randomly partitioned training and testing datasets, respectively. Initial results show that all validating models perform very close to each other. The best-trained model yields 80.67% accuracy, 86.33% precision, 81% recall, and 79.67% F1-score which shows satisfactory performance. This initial system brings the first step to a more reliable flood monitoring system.