Developing a flood image detection model using deep learning algorithms

As the impact of climate change intensifies, the frequency of short-duration heavy rainfall events gradually increases, posing a serious challenge to urban infrastructure and underground drainage systems. Assessing flood-prone areas and disaster extents relies heavily on manual surveys, lacking real-time and effective methodologies. Our study uses Mask R-CNN deep learning and closed-circuit television (CCTV) flood images to develop a real-time and effective flood detection model. The results of our study demonstrate that the proposed flood image recognition model achieves a precision of 60.6%, a recall rate of 92.2%, and an F1 score of 73.1 for the flood category. These results signify the model's exceptional capability of the model in flood detection. Additionally, through on-site measurements of road dimensions and binary matrix-based area estimation, the average error is only 1.6%. This model can be applied effectively and serves as a reference for authorities to promptly determine the occurrence of flooding and the extent of the disaster, thus facilitating the formulation of more effective disaster response measures. The developed model exhibits promising potential for real-time flood detection in urban disaster management, providing a valuable tool for authorities to respond promptly to the dynamic challenges posed by climate change.