Retrieving cloud ice masses from geostationary images with neural networks
Licentiatavhandling, 2023

Clouds are essential to the Earth's energy budget and atmospheric circulation. Despite this, many cloud parameters are poorly known, including the mass of frozen hydrometeors. On the one hand, there will be specialized satellite missions targeting such hydrometeors. On the other hand, existing satellite data can be leveraged. There should be a particular interest in using geostationary satellite observations since they provide continuous coverage. Traditionally, retrievals of cloud ice masses from geostationary measurements require solar reflectances, ignore any spatial correlations, and solely retrieve the vertically-integrated ice mass density, known as the ice water path.

This thesis challenges the traditional approach by applying supervised learning against CloudSat collocations, the only existing satellite mission targeting ice clouds. A set of neural networks is assembled to compare the performance of using different visible or infrared channels as retrieval input as well as the added value of using spatial context. The retrievals are probabilistic, in the sense that all neural networks predict quantiles to estimate the retrieval irreducible uncertainty, and thus represent the state of the art for atmospheric retrievals.

With several spectral channels, infrared retrievals are found to have a similar performance compared to the peak accuracy offered by the combination of visible and infrared channels. However, the infrared-only retrievals enable a consistent diurnal performance. The use of spatial information reinforces the retrievals, which is demonstrated by the ability to provide skilful three-dimensional estimates of ice masses, known as ice water content, from only one infrared channel. The latter retrieval scheme is supported by an extensive validation with independent measurements.

These neural network-based retrievals offer the possibility to derive new insights into cloud physics, reduce present ice cloud uncertainties, and validate climate models. Ideally, such retrieval schemes will complement the sparse measurements from specialized instruments. Finally, this thesis contains the groundwork for executing retrievals on multidecadal geostationary observations, offering unprecedented spatially and temporally continuous three-dimensional data for the tropics and mid-latitudes. The implementation of these ongoing retrievals is publicly released as part of the Chalmers Cloud Ice Climatology.

machine learning

geostationary satellites

ice clouds

EA, EDIT, Campus Johanneberg, Hörsalsvägen 11
Opponent: Assoc. Prof. Abhay Devasthale and Dr. Nina Håkansson, Swedish Meteorological and Hydrological Institute

Författare

Adrià Amell Tosas

Chalmers, Rymd-, geo- och miljövetenskap, Geovetenskap och fjärranalys

Ice water path retrievals from Meteosat-9 using quantile regression neural networks

Atmospheric Measurement Techniques,;Vol. 15(2022)p. 5701-5717

Artikel i vetenskaplig tidskrift

A. Amell, and S. Pfreundschuh, and P. Eriksson. The Chalmers Cloud Ice Climatology: retrieval implementation and validation

Robust Estimation of Global precipitation using Neural networks (REGN)

Rymdstyrelsen (154/19), 2020-01-01 -- 2023-12-31.

Ämneskategorier

Matematik

Data- och informationsvetenskap

Geovetenskap och miljövetenskap

Infrastruktur

C3SE (Chalmers Centre for Computational Science and Engineering)

Utgivare

Chalmers

EA, EDIT, Campus Johanneberg, Hörsalsvägen 11

Online

Opponent: Assoc. Prof. Abhay Devasthale and Dr. Nina Håkansson, Swedish Meteorological and Hydrological Institute

Mer information

Senast uppdaterat

2023-09-29