Real-time flash flood detection employing the YOLOv8 model

Abstract

Human lives and property are threatened by Flash floods (FF) worldwide and as a result of the unprecedented conditions of the climate change effects the losses are predicted to increase in the future. As it seems difficult to avoid and prevent them, real-time flash flood detections could be an appropriate solution for damage reduction and better management. Currently, the development of computer vision applications such as deep learning and AI has been advanced. Although AI models have been developed for applications in many fields, their implementations for geosciences are limited based on large amounts of training data and the highly required computational infrastructure. Hence, this work aims to train the latest YOLOv8 model and apply it to real-time flash flood detection for regions of Korea and possibly for other nations. To overcome the shortage of training data, we created small on-site flash flood models and took pictures and footage of them. More than 1500 photos of FF were used for model trains and validations gaining a model mean average precision of above 60% of all training depths (25, 50, 75, and 100 epochs). Despite some model false positives and missed false positive detections using the Korean FF test dataset, the YOLOv8 best model generated bounding boxes (BB) with high confidence values in most FF events. Furthermore, the robustness of the model is highlighted by its ability to smoothly detect the precise positions of the FF areas with high confidence values (best 0.86) when applied for input footage and webcam streams. It is highly encouraged to establish a real-time FF warning system to reduce their negative effects. Although YOLO is effective and fast, like other deep learning models, it requires large input data to ensure higher accuracy and confidence. Future works might explore this aspect, particularly the data acquired in light inefficiency to improve the model detections at night time.