Assessment of deterministic and probabilistic precipitation nowcasting techniques over New York metropolitan area

Abstract

This study evaluates spatiotemporal variability of deterministic and probabilistic precipitation nowcasting models' performance over the greater New York City area. Five deterministic (LINDA-D, ANVIL, S-PROG, RM-DR, RM-S) and two probabilistic (LINDA-P and STEPS) nowcasting methods were assessed using Multi-Radar Multi-Sensor (MRMS) data from 2014 to 2022. Three lead times of 2, 4 and 6 h were considered. LINDA-P had the best average Pearson's correlation of 0.87 at the first step and 0.47 at the last one and the longest 80 min average decorrelation times. Its Mean Absolute Error (MAE) was four times lower in winter than in summer, 0.3 mm vs. 1.2 mm, respectively. LINDA-P scored better than other Lagrangian Persistence-based approaches for moderate and heavy rainfall. Overall, LINDA algorithms outperformed the other nowcasting methods, but at twice the average runtime cost compared to STEPS. Uncertainties can be attributed, among others, to the models' capability to simulate the dissipation and growth processes.