Nowcasting Rainfall in Sri Lanka using Commercial Microwave Links

Abstract

Disasters like inland floods and landslides are a cause of extreme rainfall. To increase the time to take early measures against such disasters it is of great importance to have access to accurate prediction of the rainfall. For the prediction of floods, Quantitative Precipitation Forecasts (QPFs) are used as input for hydrologic models. Numerical Weather Prediction (NWP) models are commonly used to generate QPFs, but for short lead times (less than 6 hours) the NWP forecasts are not accurate enough. For very short-term forecasting the nowcasting method is used. Nowcasting rainfall is a computational process of extrapolating the most recent rainfall observations and has great potential for lead times up to 6 hours. A source for opportunistic sensing to generate rainfall estimations is Commercial Microwave Links (CMLs). Signals for telecommunication purposes will travel along a CML (link) path from one station to another. When a rainfall event occurs, the signal attenuates. This attenuation can be used for the estimation of path-averaged rainfall intensity estimations. This thesis investigates the possibilities of using CMLs for estimating and nowcasting the rainfall. The study area is Sri Lanka, a country without access to radar-based rainfall estimation. A 3.5-month CMLs data set with 2560 unique links from Dialog Sri Lanka is used and compared to 8 hourly and 12 daily rain gauge stations. The studied period is from September 12 to December 31 in 2019. To generate rainfall estimations from the CMLs, the algorithm RAINLINK is used. RAINLINK includes a set of default parameters, which was optimized for the Netherlands. For Sri Lanka, two new optimal parameters are derived. For the optimal parameters daily CML estimations are calibrated with the 12 daily rain gauges. With the new optimal parameters, the rainfall estimations are calculated and used in Pysteps, to generate nowcasts for 20 rainfall events in the studied period. Deterministic and probabilistic nowcasts are generated. These nowcasts are compared with a benchmark nowcast, called Eulerian Persistance (EP). RAINLINK gives more accurate rainfall estimations in Sri Lanka with the new parameters of the wet antenna attenuation and the coefficient compared to the default parameters. Longer path lengths tend to have lower wet antenna attenuation and a higher value for the coefficient. Nowcasts were made for the 20 selected rainfall events. The accuracy was calculated with two verification methods: the Fraction Skill Score and the Critical Success Index. From these results, the deterministic nowcasts with S-PROG and benchmark EP are accurate for lead times up to an hour. The deterministic nowcasts give the most accurate results, compared to the EP and the probabilistic nowcasts. The probabilistic nowcasts show the least accurate results, and are with a lead time of one hour only accurate for low rainfall intensities (1 mm/hr). Overall, the nowcasting results are showing that with CML rainfall estimations, Pysteps can make accurate nowcasts in Sri Lanka. To improve the nowcasts even more, a combination of rainfall estimations from CMLs, rain gauges, and satellite data could be considered.