A river flooding detection system based on deep learning and computer vision

Although floods cause millions of dollars in economic and social losses each year, many people living in developing countries, such as Brazil, do not have access to a flooding alert system because of its cost. To address this issue, we propose a cheap and robust River Flooding Detection System, which can be easily deployed in any river with a flat surface at its bedside. The novelty of our system is the use of raw images from off-the-shelf cameras with no preprocessing required. Hence, our methodology can be deployed using existing surveillance cameras in urban environments. The proposed system measures the river level by first performing a semantic segmentation of the river water blade using Deep Neural Networks (DNNs). Then, it uses Computer Vision (CV) to estimate the water level. If the water level is near or above a dangerous threshold, it sends alerts automatically without human intervention. Moreover, our system can successfully measure a river's water level with a Mean Absolute Error (MAE) of 3.32 cm, which is enough to detect when a river is about to overflow. The system is also reliable in measuring a river's water level from different camera viewpoints and lighting conditions. We show our approach's viability and evaluate our prototype's performance and overhead by deploying it to monitor two urban rivers in the city of Sao Carlos, SP, Brazil.

Automated summary

Floods cause significant economic and social losses, but people in developing countries often lack access to expensive flooding alert systems. To address this, the authors propose a cheap and robust River Flooding Detection System that utilizes raw images from off-the-shelf cameras without preprocessing. The system accurately measures river levels using semantic segmentation and computer vision techniques, and automatically sends alerts when water levels reach or exceed dangerous thresholds.