Near-infrared surveillance video-based rain gauge

Widespread surveillance cameras that continuously record rainfall information offer an opportunity for high-spatiotemporal resolution rainfall inversion. Surveillance video-based rainfall data estimation has become one of the most promising methods. However, existing relevant studies have focused on rainfall observations during the day. Little attention has been given to developing a surveillance camera-based rain gauge that works at night. Generally, ordinary surveillance cameras actively emit near-infrared light (NIR) to supplement the insufficient illumination of a surveillance scenario, providing an essential prerequisite for rainfall observations at night. In this paper, an NIR-surveillance video-based rain gauge (NIR-VRG) was constructed. First, combining the meteorological and microphysical characteristics of raindrops with the camera imaging principle, the abilities of different NIR-surveillance cameras to capture raindrops during different rainfall scenarios were discussed, providing the theoretical basis for subsequent work; second, a tensor-based algorithm was proposed for rain streak extraction from NIR surveillance video; and finally, a one-dimensional convolutional neural network (1D CNN)-based regression algorithm was proposed and used to build a mapping relationship between the extracted rain streaks and the rainfall intensity (RI). Experimental results on synthetic rainy videos showed that the proposed rain streak extraction algorithm achieves robust performance in a light breeze (speed approximately 3 m/s) and still works in a gentle breeze (speed approximately 5 m/s). Moreover, experiments during various rainfall scenarios show that the designed NIR-VRG measures rainfall information with high accuracy. The relative error of the RI and cumulative rainfall ranged from 8.86 % to 84.84 % and 7.82 % to 30.70 % respectively. The NIR-VRG fills the gap of rainfall observations at night by using surveillance cameras and provides a reference for constructing an all-weather rainfall observation network.