Impact of Radar Data Assimilation on Regional Precipitation Forecasts of Two Landfalling Atmospheric Rivers

August 27, 2026

The paper titled “Impact of Radar Data Assimilation on Regional Precipitation Forecasts of Two Landfalling Atmospheric Rivers” was recently published in the Monthly Weather Review. This work is led by CW3E’s Jia Wang, and co-authored by Minghua Zheng, Jonathan Rutz, Luca Delle Monache, Julie Kalansky, and Fred Martin Ralph. This study was supported by Atmospheric River Program Phase III and IV awarded by California Department of Water Resources (contract: 4600014294, 4600014942), and a National Oceanic and Atmospheric Administration award (NA20OAR4320278).

Radar observations are invaluable for monitoring and predicting high-impact weather and have been widely assimilated to improve short-range forecasts of warm-season extremes such convective severe weather and hurricanes. However, their assimilation for winter weather remains relatively limited in the literature. This study investigates the impacts of radar data assimilation on precipitation forecasts during two high-impact landfalling atmospheric river events in January and October 2021. This work supports CW3E’s priority, identified in the the 2025-2029 CW3E Strategic Plan, of using novel observations to improve atmospheric river related precipitation forecasts. The assimilation system uses the Gridpoint Statistical Interpolation (GSI) three-dimensional ensemble-variational method. We modified the community GSI to enable direct reflectivity assimilation by adding reflectivity as a control variable. We also revised the radial velocity thinning procedure to generate super-observations at a horizontal resolution comparable to the model grid spacing. Radial velocity and reflectivity observations are obtained from the Next Generation Weather Radar level II dataset and the Multi-Radar Multi-Sensor product, respectively.

Data impacts are evaluated through data denial experiments using a cycling framework with a 30-min assimilation frequency over a 6-h window, followed by a 6-h free forecast. Assimilation of radial velocity yields modest improvements in the precipitation forecasts during cycling in both cases, with post-cycling benefits observed mainly in the more dynamically driven October case (Fig. 1). Reflectivity assimilation generally produces larger, but more variable, improvements. In both cases, benefits persist through cycling and for 3–4 hours afterwards in the late-window experiments during the heavy precipitation stage, whereas mixed impacts are observed in the early-window experiments during precipitation onset. Assimilating both datasets largely mirrors reflectivity-only responses while retaining some benefits from radial velocity.

Most forecasts show improvement, while a smaller fraction show degradation. To understand the source of degradation, sensitivity experiments assimilating positive and negative reflectivity innovations (i.e., observation minus model background) separately show that forecast degradation is primarily associated with negative innovations (Fig. 2). Further analysis indicates that these innovations are partly contaminated with representativeness errors caused by vertical-scale mismatches between radar sampling volumes and model grids when radar beam geometry is neglected in the reflectivity forward operator (Fig. 3). These findings highlight the importance of accounting for beam geometry in the forward operator to reduce representativeness errors and increase the benefits of radar data assimilation.

Figure 1. Differences in Gilbert skill scores (GSSs) verified against gauge measurements: (a,d,g,j) withRadar_V – noRadar, (b,e,h,k) withRadar_Z – noRadar, and (c,f,i,l) withRadar_VZ – withRadar_Z. Rows correspond to Jan-Early (first), Jan-Late (second), Oct-Early (third), and Oct-Late (fourth). Gray horizontal lines separate forecasts within the cycling window (lower half) from free forecasts (upper half). Scores are masked in gray where the precipitation base rate <5%. (Figure 8 from Wang et al. (2026))

Figure 2. Differences in Gilbert skill scores (GSSs) using Pass 2 as the verification dataset: (a,b) withRadar_ZnegOmB – noRadar, (c,d) withRadar_ZposOmB – noRadar, and (e,f) withRadar_ZposOmB – withRadar_Z. Columns correspond to Jan-Early (left) and Oct-Early (right). (Figure 13 from Wang et al. (2026))

Figure 3. Medians of reflectivity innovations (i.e., observed minus background reflectivity; dBZ) for (a) the January case (0000 UTC 27 to 0600 UTC 28 January 2021) and (b) the October case (0600 UTC 24 to 1200 UTC 25 October 2021), using hourly data on WRF horizontal grids and MRMS vertical levels. Negative reflectivity values are treated as zero. Samples are included only when both observed and simulated precipitation rates are no less than 0.01 mm h^–1, and when a freezing level (0°C height in WRF simulations) exists. Gray contours (interval: 5000) represent sample counts. (Figure 14 from Wang et al. (2026))

Citation:

Wang, J., Zheng, M., Rutz, J., Delle Monache, L., Kalansky, J., & Ralph, F. M. (2026). Impact of Radar Data Assimilation on Regional Precipitation Forecasts of Two Landfalling Atmospheric Rivers. Monthly Weather Review, 154(8), 1657-1683. https://doi.org/10.1175/MWR-D-25-0222.1