Objectives Precipitable water vapor (PWV) plays an increasingly significant role in the quantitative study of the potential meteorological factors that cause rainfall. The PWV-based three-factor (PWV, PWV change and rate of PWV change) threshold method for the rain forecast has been established, empirically proving its effectiveness in some scenarios. However, an apparent issue is that not fully using real-time information restricts performance. We propose an improved monthly threshold method to tackle this problem.

Methods The basic idea is to refine the predictors. The three-factor monthly threshold method based on PWV, PWV increment and rate of PWV increment is used to realize the short-term rainfall forecast within 6 h, in which PWV is used as the main predictor, and PWV increment and PWV increase rate are used as secondary factors to assist in the prediction. We offer a quantitative standard for picking the threshold based on maximizing critical success index (CSI).The specific prediction steps are as follows: (1) The time series of the three predictors are extracted from the raw data using a 6 hours sliding window. Within each window, the PWV increment and the rate of PWV increment are computed. The window is then shifted by 30 minutes, and the process is repeated, resulting in sequences for both the PWV increment and the rate of PWV increment. (2) The monthly threshold is determined using data from 2015 to 2018, with the optimal PWV threshold for each month established based on the CSI. Once the threshold is determined, the CSI maximum principle is applied to simultaneously select the optimal thresholds for both the PWV increment and the rate of PWV increment. (3) The optimal thresholds, determined by the 2019 data, are subsequently applied to predict rainfall over the next 6 hours. Rainfall is predicted if the PWV exceeds the threshold at any given time. In cases where the PWV does not exceed the threshold, but both the PWV increment and the rate of PWV increment surpass their respective thresholds, rainfall is also predicted. If neither condition is met, no rainfall is predicted. The predicted rainfall is then compared with the observed rainfall to calculate key performance metrics, including the correct rate (CR), false alarm rate (FAR), and CSI. Additionally, it is argued that, unlike the probability of detection (POD), the use of CR provides a more reasonable evaluation of the proportion of correctly predicted rainfall events.

Results We apply our method to 11 different stations of SuomiNet. The CR of the improved method is above 89%, while the FAR is controlled below 73%. Among them the CR of the three stations P031, P047 and CN00 is about 95%, and the FAR is not higher than 65%. In general, the proposed method can predict more than 92% of rainfall on average, and the average FAR of each forecast is about 63%. Compared with the traditional three-factor algorithm, the average CR of the improved three-factor algorithm is increased by nearly 6%, the average FAR is reduced by more than 4%, and the average POD is at the same level.

Conclusions Using PWV increment and the rate of PWV increment, the new predictor can better reflect the characteristics of the rising phase of PWV and keep updating synchronously with PWV. Compared with the traditional method, our algorithm can predict rainfall more effectively and has higher applicability.