Obtaining 3D convective characteristics from a machine learning-based integration of multi-sensor satellite observations
Loading...
Date issued
Authors
Editors
Journal Title
Journal ISSN
Volume Title
Publisher
Reuse License
Abstract
Convective clouds are key players in the climate system of the Earth. They influence both in-
coming solar and outgoing terrestrial radiation, and they regulate the hydrological cycle through
complex feedback mechanisms. Despite their importance, clouds remain one of the largest sources
of uncertainty in climate models — posing persistent challenges to scientists around the globe.
Among all cloud types, convective systems stand out due to their ability to evolve rapidly from
harmless cumulus clouds into intense thunderstorms. Accurately predicting this evolution is vital
for effective risk assessment and mitigation — especially as extreme weather events are expected
to occur more frequently in a warming world.
Satellite observations offer profound insights into the behavior of convective clouds. Although the
volume of satellite data has grown tremendously in recent decades, extracting meaningful patterns
from these large and complex datasets remains a daunting task. However, recent advances in
machine learning have introduced new tools that can help address this challenge. While satellite
data often lack fine vertical resolution or have limited temporal coverage, machine learning tech-
niques allow to bridge these gaps, revealing previously hidden patterns associated to the dynamic
evolution of cloud systems.
This thesis addresses current challenges by developing a machine learning framework that com-
bines multiple satellite datasets to improve our understanding of convective clouds. Specifically,
it uses 2D imagery from the geostationary MSG SEVIRI satellite to predict 3D cloud structures
as observed by the CloudSat cloud profiling radar, which provides 2D vertical cross-sections of
the radar reflectivity. This approach helps overcome current limitations in vertical cloud profiling.
The model evaluation demonstrates that it can accurately reconstruct both the vertical structure
and the distribution of hydrometeors. By leveraging the high spatial and temporal resolution of
MSG SEVIRI alongside the vertical detail from CloudSat, this method considerably enhances the
availability of 3D cloud structures across broad regions on Earth.
Building on this foundation, the thesis applies the predicted 3D cloud fields to investigate tropical
convective cloud behavior, with a focus on the role of convective cores and large-scale spatial
clustering in shaping cloud structure. An adapted, object-based detection algorithm first identifies
individual cloud objects and their cores, then tracks their evolution over time. This method
may provide detailed insights into how cores influence the 3D structure of clouds and how such
patterns relate to broader phenomena — such as convective organisation, a process closely tied to
extreme weather events. The findings also reveal connections between cloud morphology, convective
clustering, and larger-scale atmospheric dynamics, including the seasonal migration of the ITCZ.
Although this work centers on convective clouds, the machine learning framework developed here
has broader applicability for questions related to atmospheric and climate sciences. By improving
access to high-resolution, 3D atmospheric structures, it offers valuable tools for a wide range of
studies on cloud processes — and lays the groundwork for supporting a more accurate assessment
of climate-related risks in the future.