Development of cloud classes for machine learning scene classification in the context of the FORUM mission

Sildatke, Sarah (2026) Development of cloud classes for machine learning scene classification in the context of the FORUM mission. [Laurea magistrale], Università di Bologna, Corso di Studio in Science of climate [LM-DM270]
Documenti full-text disponibili:
[thumbnail of Thesis] Documento PDF (Thesis)
Disponibile con Licenza: Creative Commons: Attribuzione - Non commerciale - Condividi allo stesso modo 4.0 (CC BY-NC-SA 4.0)

Download (16MB)

Abstract

The cloud response to warming remains one of the largest uncertainties in climate science. The Far-infrared Outgoing Radiation Understanding and Monitoring (FORUM) mission will provide the first comprehensive spectral measurements of the far-infrared, where a large fraction of the outgoing longwave radiation is emitted. Interpreting these measurements requires a reliable scene classification, yet established cloud types are defined by appearance rather than by physical properties. This thesis tests whether unsupervised machine learning can derive physically meaningful cloud classes from atmospheric profiles, and whether these classes can train the Scene and Cloud Identification and Classification (SCIC) algorithm on FORUM-like spectra. CAMS model profiles are divided into regions and separated into clear-sky, liquid-only and ice-cloud scenes. The ice-cloud profiles are clustered with Self-Organising Maps and grouped into four physically ordered classes per region. Simulated FORUM-like spectra of the same profiles are used to train and test SCIC in three representative regions. The ice-cloud classes show a similar structure in climatically different regions and agree with the cloud levels of the World Meteorological Organization. Optically thin and thick high ice clouds are identified reliably, with precisions of 78-91% in the tropical and mid-latitude regions, whereas the lower classes and liquid-only scenes are often confused. The polar region is separated least well. Merging the confused lower classes in one mid-latitude region raises the balanced accuracy from 71% to 79%. Data-driven classes can therefore provide a usable SCIC training set for high ice clouds, while the number of useful classes is limited by spectral separability rather than by physical distinctness.

Abstract
Tipologia del documento
Tesi di laurea (Laurea magistrale)
Autore della tesi
Sildatke, Sarah
Relatore della tesi
Correlatore della tesi
Scuola
Corso di studio
Ordinamento Cds
DM270
Parole chiave
FORUM,far infrared,cloud classification,self-organizing map,SCIC,machine learning
Data di discussione della Tesi
28 Ottobre 2026
URI

Altri metadati

Statistica sui download

Gestione del documento: Visualizza il documento

^