Understanding tropical cyclones using machine learning with satellite imagery
File(s)
Author(s)
Smith, Mohan
Type
Thesis
Abstract
Large scale (0 ≤ r ≤ 2200 km) and small scale (0 ≤ r ≤ 550 km) infra red satellite imagery
of tropical cyclones is examined using a variety of machine learning techniques with varying
degrees of complexity. Regarding large scale images, a dipole in outgoing long-wave radiation
is revealed through composite analysis. The dipole is composed of low fluxes within the
tropical cyclone circulation (r < 500 km) with high fluxes offset (1000 - 2000km) north
(poleward) west of the tropical cyclone center. Both the strength and orientation of the
dipole are independent of vertical wind shear and cyclone intensity. Using reanalysis data, it
is shown that divergence at surface and convergence aloft facilitates subsidence throughout
the troposphere over the region of high fluxes. The divergent signal at surface is produced
over a region where the cyclone pressure gradient competes with the background meridional
pressure gradient, splitting the environmental winds. Aloft the pressure gradients are reversed
and environmental winds are impeded by the cyclone outflow producing a convergent signal.
Subsidence is induced in both the cyclone outflow and the environmental wind over this
region. This is a region, within r ≤ 2200 km from the cyclone center, that consistently
accommodates the subsidence of the secondary circulation.
Using the ECMWF (European Center for Mid-range Weather Forecasting) ensemble, a
significant relationship is identified between the dipole orientation and the cyclone path. The
high flux region is positioned north west of the cyclone center for cyclones propagating westward.
The high flux region is positioned west of the cyclone center for cyclones propagating
eastward. Different environmental flows and pressure regimes, between each case, can explain
the anti-clockwise dipole rotation as cyclones recurve towards the east. In this way
the dipole configuration is a reflection of some of the vortex environmental interaction that
also determines the cyclone track. It is demonstrated that a sub-ensemble can be formed
by selecting ensemble members with similar outgoing long-wave radiation (OLR) fields. The
strongest improvements (8%) in 24 hour track forecasts are found when sub-ensembles are selected by OLR fields that encompass the dipole. Ensemble members with similar dipole
properties demonstrate matching environment interactions, and subsequently, coherent future
track paths. Through comparison with other regions of interest in large scale OLR fields,
it is demonstrated that the high OLR component of the dipole bears a strong association
with future TC tracks. The reductions in track forecast error using sub-ensembles selected by
OLR fields are similar to the error reductions found when selecting sub-ensembles by cyclone
position. Both of these techniques require accurate cyclone positions.
A novel technique to identify the center of tropical storms and tropical cyclones is presented.
It is shown that tropical storms demonstrate a highly variable, elliptical, asymmetric
cloud structure, typical of mesoscale convective systems. Sequential long and short wave
infra red images, encompassing the region r ≤ 550 km, are used to create cyclone videos.
Neural networks incorporating both convolutional, and recurrent layers, are trained upon the
cyclone videos to identify the cyclone center. A 40% reduction in position error is provided
across all intensities over the baseline ARCHER-2 product. Of the networks tested, the best
performance was found with networks that initially identify spatial features within each image,
and then subsequently identify the temporal evolution of such features. Long-wave infra
red channels provide a 25% reduction in position error over short-wave infra red channels.
It is proposed that the capacity for long-wave infra red channels to image throughout the
troposphere in clear-sky conditions provides this error reduction.
2
of tropical cyclones is examined using a variety of machine learning techniques with varying
degrees of complexity. Regarding large scale images, a dipole in outgoing long-wave radiation
is revealed through composite analysis. The dipole is composed of low fluxes within the
tropical cyclone circulation (r < 500 km) with high fluxes offset (1000 - 2000km) north
(poleward) west of the tropical cyclone center. Both the strength and orientation of the
dipole are independent of vertical wind shear and cyclone intensity. Using reanalysis data, it
is shown that divergence at surface and convergence aloft facilitates subsidence throughout
the troposphere over the region of high fluxes. The divergent signal at surface is produced
over a region where the cyclone pressure gradient competes with the background meridional
pressure gradient, splitting the environmental winds. Aloft the pressure gradients are reversed
and environmental winds are impeded by the cyclone outflow producing a convergent signal.
Subsidence is induced in both the cyclone outflow and the environmental wind over this
region. This is a region, within r ≤ 2200 km from the cyclone center, that consistently
accommodates the subsidence of the secondary circulation.
Using the ECMWF (European Center for Mid-range Weather Forecasting) ensemble, a
significant relationship is identified between the dipole orientation and the cyclone path. The
high flux region is positioned north west of the cyclone center for cyclones propagating westward.
The high flux region is positioned west of the cyclone center for cyclones propagating
eastward. Different environmental flows and pressure regimes, between each case, can explain
the anti-clockwise dipole rotation as cyclones recurve towards the east. In this way
the dipole configuration is a reflection of some of the vortex environmental interaction that
also determines the cyclone track. It is demonstrated that a sub-ensemble can be formed
by selecting ensemble members with similar outgoing long-wave radiation (OLR) fields. The
strongest improvements (8%) in 24 hour track forecasts are found when sub-ensembles are selected by OLR fields that encompass the dipole. Ensemble members with similar dipole
properties demonstrate matching environment interactions, and subsequently, coherent future
track paths. Through comparison with other regions of interest in large scale OLR fields,
it is demonstrated that the high OLR component of the dipole bears a strong association
with future TC tracks. The reductions in track forecast error using sub-ensembles selected by
OLR fields are similar to the error reductions found when selecting sub-ensembles by cyclone
position. Both of these techniques require accurate cyclone positions.
A novel technique to identify the center of tropical storms and tropical cyclones is presented.
It is shown that tropical storms demonstrate a highly variable, elliptical, asymmetric
cloud structure, typical of mesoscale convective systems. Sequential long and short wave
infra red images, encompassing the region r ≤ 550 km, are used to create cyclone videos.
Neural networks incorporating both convolutional, and recurrent layers, are trained upon the
cyclone videos to identify the cyclone center. A 40% reduction in position error is provided
across all intensities over the baseline ARCHER-2 product. Of the networks tested, the best
performance was found with networks that initially identify spatial features within each image,
and then subsequently identify the temporal evolution of such features. Long-wave infra
red channels provide a 25% reduction in position error over short-wave infra red channels.
It is proposed that the capacity for long-wave infra red channels to image throughout the
troposphere in clear-sky conditions provides this error reduction.
2
Version
Open Access
Date Issued
2021-01
Date Awarded
2021-05
Copyright Statement
Creative Commons Attribution NonCommercial NoDerivatives Licence
Advisor
Toumi, Ralf
Sponsor
Natural Environment Research Council (Great Britain)
Grant Number
NE/L002515/1
Publisher Department
Physics
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)