KU River Dataset for Waterbody Segmentation in South Korea: Application of Foundation Model with Auto-prompting

ABSTRACT

The increasing frequency of extreme rainfall caused by climate change raises flood risks in Korea, provoking demands for new, effective methods for early flood detection. Recently, deep learning-based waterbody segmentation methods using surveillance camera images have become a focal point for effective flood detection. Many studies have developed waterbody segmentation models using benchmark datasets collected from diverse regions that represent their unique geographical and environmental characteristics. While benchmark datasets offer valuable insights, they often lack sufficient data to generalize environmental features on a universal scale. To handle the limitations of the generalization, models should learn regional features of the applied environment to ensure reliable flood detection performance. Although the need for developing region-specific datasets designed for local application has risen, little efforts have been devoted to constructing Korean river datasets due to the lack of resources and time. To address these challenges, this study introduces a novel region-specific waterbody segmentation dataset called KU River Dataset and proposes automated prompting, an advanced model training technique. First, KU River Dataset consists of 280 river and stream images and is specifically designed to reflect the diverse characteristics of the environment in Korea under various light conditions and surrounding landscapes. Second, an automated prompting method adapting a foundation model enhances the model’s performance using limited data. We employed Segment Anything Model 2 (SAM 2), a foundation model for image segmentation tasks. The automated prompts, generated from SAM 2’s image encoder, guide the model to focus on features of the waterbody. As a result, SAM 2 trained with KU River Dataset achieved 5% improvement in Intersection over Union (IoU) score on the test set compared to SAM 2 trained with a benchmark dataset of the same size. These results demonstrate the effectiveness of applying a region-specific dataset and an automated prompting method for improving regional flood detection. To improve the model’s robustness across diverse environmental conditions, including low-light and flood scenarios, we plan to gather more images of night vision and inundated riversides. Through the further development of our dataset, we expect to enhance the precision of early flood detection systems.

ACKNOWLEDGEMENT

This work was supported by Korea Environment Industry & Technology Institute (KEITI) through R&D Program for Innovative Flood Protection Technologies against Climate Crisis, funded by Korea Ministry of Environment (MOE)(RS-2023-00218873).