Leveraging a radar-based disdrometer network to develop a probabilistic precipitation phase model in eastern Canada

Abstract

This study presents a probabilistic model that partitions the precipitation phase based on hourly measurements from a network of radar-based disdrometers in eastern Canada. The network consists of 27 meteorological stations located in a boreal climate for the years 2020-2023. Precipitation phase observations showed a 2 m air temperature interval between 0-4 °C, where probabilities of occurrence of solid, liquid, or mixed precipitation significantly overlapped. Single-phase precipitation was found to occur more frequently than mixed-phase precipitation. Probabilistic phase-guided partitioning (PGP) models of increasing complexity using random forest algorithms were developed. The PGP models classified the precipitation phase and partitioned the precipitation accordingly into solid and liquid amounts. PGP-basic is based on 2 m air temperature and site elevation, while PGP-hydromet integrates relative humidity, surface pressure, and precipitation rate. PGP-full includes all previous data, along with atmospheric reanalysis data, the 1000-850 hPa layer thickness, and temperature lapse rate. The PGP models were compared to benchmark precipitation-phase-partitioning methods. These included a model with a single temperature threshold set at 1.5 °C, a linear-transition model with dual temperature thresholds of -0.38 and 5 °C, and a psychrometric balance model. Among the benchmark models, the single temperature threshold had the best classification performance (F1 score of 0.74) due to a low count of mixed-phase events. The other benchmark models tended to over-predict mixed-phase precipitation in order to decrease the partitioning error. All PGP models showed significant phase classification improvement by reproducing the observed overlapping precipitation phases based on 2 m air temperature. PGP-hydromet and PGP-full displayed the best classification performance (F1 score of 0.84). In terms of partitioning error, PGP-full had the lowest RMSE (0.27 mm) and the least variability in performance. The RMSE of the single-temperature-threshold model was the highest (0.40 mm) and showed the greatest performance variability. An input variable importance analysis revealed that the additional data used in the more complex PGP models mainly improved mixed-phase precipitation prediction. The improvement of mixed-phase prediction remains a challenge. Relative humidity was deemed to be the least important input variable used due to consistent near-saturation water vapour conditions. Additionally, the reanalysis atmospheric data proved to be an important factor in increasing the robustness of the partitioning process. This study establishes a basis for integrating automated phase observations into a hydrometeorological observation network and for developing probabilistic precipitation phase models.