Two-Step Machine Learning Process for Precipitation Nowcasting over the UAE Region

Nowcasting refers to forecasting over a short time period, typically 0 to 3 hours. Extreme precipitation has been occurring more frequently throughout the world recently. Accurate nowcasting and the precise Quantitative Precipitation Estimation (QPE) are important for disaster mitigation applications due to floods. Precipitation nowcasting has been widely studied. However, recent research on radar-based precipitation nowcasting often uses theoretical equations to estimate rainfall rates from radar observations. The precipitation nowcasting is studied using this estimated rainfall. However, the theoretical relationship between radar observations and rainfall presents challenges in estimating extreme precipitation. This paper proposes two-step radar-based precipitation nowcasting approach in United Arab Emirates (UAE). The UAE is a challenging region for precipitation nowcasting due to sporadic extreme precipitation events. For accurate precipitation nowcasting, two challenges need to be addressed. First, the estimation of precipitation needs to be accurate. Previous research has shown that machine learning based QPE outperforms theoretical Z-R relation [1]. Second, the movement of precipitation cells needs to be accurately predicted. In this paper, we present a two-step machine learning precipitation nowcasting to address these challenges by sequentially combining machine learning based QPE and machine learning based nowcasting. The rainfall rate is estimated from radar observations using a machine learning model. C-band vertical profiles of dual polarization radar observations were used to train machine learning models, with rain gauge measurements using as a ground reference. The QPE product from the machine learning model was then used to train the machine learning based precipitation nowcasting. The combination of a convolution neural network and a recurrent neural network was used to capture spatial and temporal variations of precipitation. The performance of the two-steps machine learning model was evaluated both quantitatively and qualitatively for accurate precipitation estimation and accurate forecasting of precipitation cells.

References

[1] EunYeol Kim, V.Chandrasekar, Chandrasekar Radhakrishnan, Luca Delle Monache, and Duncan Axisa, “Radar Based Quantitative Precipitation Estimation Using Deep Learning over the UAE region”, 104th American Meteorological Society Annual Meeting, 2024