Application of optical flow technique to short-term rainfall forecast for some synoptic patterns in Vietnam

Abstract

The occurrence of heavy rains can lead to human, economic, and ecological disasters with large-scale consequences. There are now many precipitation forecasting systems that use radar products with different algorithms and techniques to provide forecasts for up to one to three hours, such as McGill algorithm for precipitation nowcasting by Lagrange extrapolation (MAPLE), Short-Term Ensemble Prediction System (STEPS), and Short-range Warning of Intense Rainstorms in Localized Systems (SWIRLS). Optical flow engineering is an important technique in computer vision. Our aim was to apply optical flow techniques using the methods of DenseRotation_FCAFlow - this method is a small branch of the Rainymotion library system, belonging to the Dense method group DenseRotation_Farneback - this method uses the local optical flow and polynomial function extension techniques to calculate radar echo region motion, Real-time Optical Flow by Variational Methods for Echoes of Radar - this method is used with two sets of parameters for Hong Kong (ROVER_HKO) and Vietnam ( ROVER_VN) to predict rainfall quantitatively from 2019–2021 radar rainfall quantitative data in Vietnam. The results show the following: (i) Changing the parameterizations of the ROVER_HKO method to apply to conditions in Vietnam (ROVER_VN) gives better results than the remaining methods for the total accumulated rainfall of the entire rainfall events, and the ROVER_VN method gives the best results in case rain occurs due to a combination of many synoptic patterns; (ii) The performance of Quantitative Precipitation Forecasting (QPF) using a performance diagram with light and moderate rainfall thresholds is also better captured in a combined local and global optical flow method than when using either type of optical flow alone. The ROVER_VN method quantitatively forecasts the cumulative rainfall of the entire rain event in case the rain is caused by the interaction of many different synoptic patterns and the rainfall area develops to be wide and fast moving, playing an especially important role in providing input data for forecasting and warning of floods, flash floods, and landslides throughout the territory of Vietnam.

Application of Radar-Based Precipitation Data Improves the Effectiveness of Urban Inundation Forecasting

“…However, there are also limitations in the cases where the rainfall at the station at that time reaches its peak because in these cases, the rainfall tends to be estimated as higher than the actual observation. This leads to forecast results that tend to be higher than the actual occurrence [31]. Most case studies show that because the QPF method uses the ROVER_VN optical flow technique, in which the forecast uses the backward semi-Lagrangian extrapolation method [31], the forecasted field does not fully describe the development and weakening process of rainy areas, leading to a tendency for the forecasted precipitation field to deviate greatly from reality during extreme rainfall and lead-time long-term forecasting [53][54][55][56][57][58].…”

“…Detailed improvements were shown by Kimpara et al [15]. The method of calculating QPF from radar uses the optical flow technique with a set of adjustable parameters, such as the ROVER_VN model described by Thu et al [31]. QPF precipitation is also updated continuously every 10 min.…”

“…QPE and QPF in Event 3 (Figure 10c) are also closely simulated with the observed values, with the only exception being at 01:00 UTC on 12 August 2022, when QPF is much higher than the observed value. The QPF error may be due to the ROVER_VN optical flow method [31] not capturing the growth or decay process of short-duration rain areas, which is also a limitation of the optical flow method that some previous studies have mentioned [24], or because the process of estimating the motion of precipitation fields is slower or faster than reality. This can also be seen in Figure 10c (or Figure 11c), in which the predicted QPF values are too high compared to observation at 01:00 UTC on 12 August 2022, while the observed data are too low (Figure 10c) or reach a maximum (Figure 11c) that occurred 1 h before (at 00:00 UTC on 12 August 2022).…”

“…With strengths in tracking continuously moving objects over a range of short periods, the optical flow technique has been applied in QPF. It has been proven to be one of the most effective techniques for extremely short-term rainfall forecasting and is currently being used in Vietnam [31]. Flood forecasting for rivers and streams often uses QPE instead of QPF because QPF has more potential than QPE [32].…”