skip to main content

Title: Machine Learning–Based Hurricane Wind Reconstruction

Here we present a machine learning–based wind reconstruction model. The model reconstructs hurricane surface winds with XGBoost, which is a decision-tree-based ensemble predictive algorithm. The model treats the symmetric and asymmetric wind fields separately. The symmetric wind field is approximated by a parametric wind profile model and two Bessel function series. The asymmetric field, accounting for asymmetries induced by the storm and its ambient environment, is represented using a small number of Laplacian eigenfunctions. The coefficients associated with Bessel functions and eigenfunctions are predicted by XGBoost based on storm and environmental features taken from NHC best-track and ERA-Interim data, respectively. We use HWIND for the observed wind fields. Three parametric wind profile models are tested in the symmetric wind model. The wind reconstruction model’s performance is insensitive to the choice of the profile model because the Bessel function series correct biases of the parametric profiles. The mean square error of the reconstructed surface winds is smaller than the climatological variance, indicating skillful reconstruction. Storm center location, eyewall size, and translation speed play important roles in controlling the magnitude of the leading asymmetries, while the phase of the asymmetries is mainly affected by storm translation direction. Vertical wind shear impacts the asymmetry phase to a lesser degree. Intended applications of this model include assessing hurricane risk using synthetic storm event sets generated by statistical–dynamical downscaling hurricane models.

more » « less
Award ID(s):
Author(s) / Creator(s):
 ;  ;  ;  ;  
Publisher / Repository:
American Meteorological Society
Date Published:
Journal Name:
Weather and Forecasting
Page Range / eLocation ID:
p. 477-493
Medium: X
Sponsoring Org:
National Science Foundation
More Like this
  1. Abstract

    The hazards induced by tropical cyclones (TCs), for example, high winds, extreme precipitation, and storm tides, are closely related to the TC surface wind field. Parametric models for TC surface wind distribution have been widely used for hazards and risk analysis due to their simplicity and efficiency in application. Here we revisit the parametric modeling of TC wind fields, including the symmetrical and asymmetrical components, and its applications in storm tide modeling in the North Atlantic. The asymmetrical wind field has been related to TC motion and vertical wind shear; however, we find that a simple and empirical background‐wind model, based solely on a rotation and scaling of the TC motion vector, can largely capture the observed surface wind asymmetry. The implicit inclusion of the wind shear effect can be understood with the climatological relationship between the general TC motion and wind shear directions during hurricane seasons. For the symmetric wind field, the widely used Holland wind profile is chosen as a benchmark model, and we find that a physics‐based complete wind profile model connecting the inner core and outer region performs superiorly compared to a wind analysis data set. When used as wind forcing for storm tide simulations, the physics‐based complete wind profile integrated with the background‐wind asymmetry model can reproduce the observed storm tides with lower errors than the often‐used Holland model coupled with a translation‐speed‐based method.

    more » « less
  2. This article presents an azimuthally asymmetric gradient hurricane wind field model that can be coupled with hurricane-track models for engineering wind risk assessments. The model incorporates low-wavenumber asymmetries into the maximum wind intensity parameter of the Holland et al. wind field model. The amplitudes and phases of the asymmetries are parametric functions of the storm-translation speed and wind shear. Model parameters are estimated by solving a constrained, nonlinear least squares (CNLS) problem that minimizes the sum of squared residuals between wind field intensities of historical storms and model-estimated winds. There are statistically significant wavenumber-1 asymmetries in the wind field resulting from both storm translation and wind shear. Adding the translation vector to the wind field model with wavenumber-1 asymmetries further improves the model’s estimation performance. In addition, inclusion of the wavenumber-1 asymmetry resulting from translation results in a greater decrease in modeling error than does inclusion of the wavenumber-1 shear-induced asymmetry. Overall, the CNLS estimation method can handle the inherently nonlinear wind field model in a flexible manner; thus, it is well suited to capture the radial variability in the hurricane wind field’s asymmetry. The article concludes with brief remarks on how the CNLS-estimated model can be applied for simulating wind fields in a statistically generated ensemble.

    more » « less
  3. Abstract

    Polygonal eyewall asymmetries of Hurricane Michael (2018) during rapid intensification (RI) are analyzed from ground‐based single Doppler radar. Here, we present the first observational evidence of the evolving wind field of a polygonal eyewall during RI to Category 5 intensity by deducing the axisymmetric and asymmetric winds at 5‐min intervals. Spectral time decomposition of the retrieved tangential wind structure shows quantitative evidence of low (1–4) azimuthal wavenumbers with propagation speeds that are consistent with linear wave theory on a radial vorticity gradient, suggesting the presence of rapidly evolving vortex Rossby waves. Dual‐Doppler winds from the NOAA P‐3 Hurricane Hunter airborne radar provide further evidence of the three‐dimensional vortex structure that supports growth of asymmetries during RI. Both reflectivity and tangential wind fields show polygonal structure and propagate at similar speeds, suggesting a close coupling of the dynamics and the convective organization during the intensification.

    more » « less
  4. null (Ed.)
    Abstract Although global and regional dynamical models are used to predict the tracks and intensities of hurricanes over the ocean, these models are not currently used to predict the wind field and other impacts over land. This two-part study performs detailed evaluations of the near-surface, overland wind fields produced in simulations of Hurricane Wilma (2005) as it traveled across South Florida. This first part describes the production of two high-resolution simulations using the Weather Research and Forecasting (WRF) Model, using different boundary layer parameterizations available in WRF: the Mellor–Yamada–Janjić (MYJ) scheme and the Yonsei University (YSU) scheme. Initial conditions from the Global Forecasting System are manipulated with a vortex-bogusing technique to modify the initial intensity, size, and location of the cyclone. It is found possible through trial and error to successfully produce simulations using both the YSU and MYJ schemes that closely reproduce the track, intensity, and size of Wilma at landfall. For both schemes the storm size and structure also show good agreement with the wind fields diagnosed by H*WIND and the Tropical Cyclone Surface Wind Analysis. Both over water and over land, the YSU scheme has stronger winds over larger areas than does the MYJ, but the surface winds are more reduced in areas of greater surface roughness, particularly in urban areas. Both schemes produced very similar inflow angles over land and water. The overland wind fields are examined in more detail in the second part of this study. 
    more » « less
  5. Abstract

    A mobile Shared Mobile Atmospheric Research and Teaching (SMART) radar was deployed in Hurricane Harvey and coordinated with the Corpus Christi, TX, WSR‐88D radar to retrieve airflow during landfall. Aerodynamic surface roughness estimates and a logarithmic wind profile assumption were used to project the 500‐m radar‐derived maximum wind field to near the surface. The logarithmic wind assumption was justified using radiosonde soundings taken within the storm, while the radar wind estimates were validated against an array of StickNets. For the data examined here, the radar projections had root‐mean‐squared error of 3.9 m/s and a high bias of 2.3 m/s. Mesovorticies in Harvey's eyewall produced the strongest radar‐observed winds. Given the wind analysis, Harvey was, at most, a Category 3 hurricane (50–58 m/s sustained winds) at landfall. This study demonstrates the utility of integrated remote and in situ observations in deriving spatiotemporal maps of wind maxima during hurricane landfalls.

    more » « less