E-Locus - Institutional Repository of the University of Crete

Home Collections School/Department School of Sciences and Engineering Department of Chemistry Post-graduate theses

Post-graduate theses

Current Record: 3 of 642

[Add to Basket]

Identifier

000463159

Title

Implementation of machine learning algorithms for fire risk prediction in Greece

Alternative Title

Εφαρμογή αλγορίθμων μηχανικής μάθησης για την πρόβλεψη του κινδύνου πυρκαγιάς στην Ελλάδα

Author

Καλογεράκη, Ελένη Ε.

Thesis advisor

Κανακίδου, Μαρία

Reviewer

Βρεκούσης, Μιχαήλ
Banks, Andrew

Abstract

Fire prediction is an especially difficult task, as the complex interactions between the fire drivers and conditions that start a fire can be challenging. In this work, we aim to predict fire risk with the use of ML algorithms. Meteorological, topographical and satellite daily data in 1km x 1km grid resolution have been utilized as parameters to train our models. The data set that was used covered the period from 2011 to 2021. The area of study comprises part of the Balkan peninsula, the whole of Greece and a part of western Turkey. In order to reduce the parameters set used to train the model, techniques such as Mutual Information and Spearman Rank Correlation were applied. We employed Random Forest (RF) and XGBoost classifiers to address a binary classification problem (fire occurrence or no fire), training our models on data from 2011- 2019 and evaluating their performance on subsequent years of 2020 and 2021. The XGBoost model stood out for its robustness, achieving a sensitivity of 95% and specificity of 50%. Our feature reduction analysis revealed that excluding Corine Land Cover class fractions had a negligible effect on the model's output, whereas further reduction including normalized difference vegetation index (NDVI) and DEM (Digital Elevation Model) variables impacted the performance slightly. This research provides valuable insights into the application of ML algorithms for wildfire prediction. For future investigation, incorporating numerical weather prediction (NWP) forecasting data and leveraging unsupervised learning methods are proposed to delve deeper into the patterns and drivers of fire incidents.

Language

English

Subject

Supervised learning

Wildfire prediction

Επιβλεπόμενη μάθηση

Issue date

2024-03-21

Collection

School/Department--School of Sciences and Engineering--Department of Chemistry--Post-graduate theses

Type of Work--Post-graduate theses

Permanent Link