ABSTRACT

Convolution neural network (CNN) has been proved to be very useful in many research fields. However, there are still many possible applications in disaster monitoring and emergency operations. In Taiwan, CCTV also plays an important role for providing real-time situation in the field. Due to its importance, there are more than 20,000 CCTV working in Taiwan to provide realtime information. During flooding events, however, operation agency still needs manpower to check every CCTV footage and it is time and labor consuming. To address this issue, we collected 52,941 CCTV images during typhoon and storm events in 2023. From these images, we selected 3,540 images within urban area during daytime as the training dataset of the CNN model for flooding detection in CCTV image in urban areas. The developed CNN model has 2.4 million parameters to train and the training time costs about 1 hour. The trained CNN model shows a good performance with overall accuracy reaching 98 % while the accuracy for flooding detection for training and testing datasets reaches 85 %. This model may be further improved by considering two-stages CNN model and hopefully may make flooding detection using CCTV in a more efficient way.

Keywords：Convolution neural network, CCTV, Flooding in urban areas.