Author(s): Jongyun Byun; Jinwook Lee; Hyeon-Joon Kim; Jongjin Baik; Changhyun Jun

Linked Author(s): Changhyun Jun

Keywords: CCTV; Convolutional Autoencoder; Supervised Learning

Abstract: Accurate rainfall forecasting is pivotal for managing environmental vulnerabilities in complex terrain. As urban areas evolve, understanding and predicting rainfall patterns become imperative for effective hydrological planning and risk mitigation. In the era of the Internet of Things (IoT), leveraging non-traditional data sources has become increasingly attractive for enhancing the spatiotemporal resolution of existing observation network. This trend is notably evident in the field of rainfall observation and recent endeavors include innovative approaches to estimate rainfall by harnessing CCTV image. Our methodology involves the application of CCTV image collected from ground observation site and intends to predict the CCTV image at a specific time point based on the preceding group of CCTV images resolving the data insufficiency. The overall output data for the model utilized rain streaks extracted from original CCTV data. K-Nearest Neighbor (K-NN) algorithm was employed to effectively separate the background and foreground. We developed a supervised learning model by constructing an encoder-decoder network based on the convolutional autoencoder architecture. This network is designed to extract the inherent characteristics of CCTV image and its rain streak and to minimize the loss between result from image prediction model and CCTV image collected in a single frame.

DOI: https://doi.org/10.3850/iahr-hic2483430201-385

Year: 2024