Simulation of regional rainfall observation using urban surveillance camera networks

Abstract Widespread surveillance cameras show excellent potential for high spatial and temporal resolution rainfall observation. As the accuracy of surveillance camera-based rain gauges (SRGs) continues to improve, surveillance camera-based rainfall observation networks (SRN) have received increasing attention worldwide. Limited by the availability of surveillance data, there is little investigation into the performance of SRNs. In this study, a simulated SRN construction model is proposed, which employs meteorology and geography research as a priori knowledge and the state-of-the-art achievements of SRG as the basis, bringing the simulated SRNs closer to their practical performance. Regional rainfall observations from SRNs were compared with those from ground gauge network (GRN), real-time and calibrated radars. The experimental results show that (1) SRNs can significantly improve correlation with real-time and calibrated radar observations over GRN; (2) an SRN consisting of 700 SRGs whose estimation accuracy is higher than 80% or 500 SRGs whose estimation is accuracy higher than 90% could achieve a comparable precision performance to that of GRN; (3) as the number of SRGs reaches 900 (the spatial density of SRGs is about 0.47 per km 2 ), the performance of the SRNs tend to be stable. Although increasing the number of cameras helps alleviate the problem of insufficient accuracy of a single SRG during heavy/violent rainfall, excessive cameras may reduce the accuracy due to the inherent measurement errors of SRG. Therefore, developing a robust SRG filtering strategy to find the optimal number of SRGs is essential. Our research provides an important reference for researchers who are holding a sceptical view of the availability of surveillance camera rainfall networks while shedding light on building a new low-cost and high-resolution SRN based on the existing surveillance camera resources.