Exploratory analysis of wildfires in Australia and a machine learning approach for wildfire modeling in Google Earth Engine
Author
Sulova, Andrea
Term
4. term
Publication year
2020
Submitted on
2020-06-04
Pages
69
Abstract
Naturbrande forventes at blive hyppigere i takt med klimaets opvarmning. I den australske sommer 2019–2020 brændte 46 millioner acres, hvilket rejser spørgsmål om, hvordan klimatiske, miljømæssige, topografiske og sociale forhold påvirker risikoen, og hvordan vi kan forudsige og forebygge brande. Dette studie analyserer de australske brande i 2019–2020 med frit tilgængelige jordobservationsdata (satellitdata). Vi udvikler en automatiseret proces til at opbygge et træningsdatasæt i kontinental skala, som mærker lokaliteter med brand og uden brand og knytter dem til mulige årsagsfaktorer, samtidig med at beregningsomkostningerne holdes nede. Vi afprøver flere maskinlæringsmetoder—Random Forest (RF), Naïve Bayes (NB) og Classification and Regression Tree (CART). Maskinlæring er algoritmer, der lærer mønstre fra data for at lave forudsigelser. Blandt disse klarer RF‑modellen sig bedst og bruges til at vurdere, hvilke variable der bidrager mest til at forudsige brandforekomst (analyse af variabel betydning). Målet er at estimere sandsynligheden for brandforekomst på tværs af Australien og at identificere de vigtigste drivkræfter for naturbrande i sæsonen 2019–2020. Resultaterne kan understøtte mere målrettede forebyggende tiltag i brandudsatte områder ved at fokusere på de mest relevante faktorer.
Wildfires are expected to become more frequent as the climate warms. During Australia’s 2019–2020 summer, fires burned 46 million acres, raising questions about how climate, environmental, topographic, and social factors shape risk and how we can predict and prevent fires. This study analyzes the 2019–2020 Australian fires using freely available Earth observation (satellite) data. We develop an automated process to build a continent-wide training dataset that labels locations with fire and no-fire outcomes and links them to potential causal factors, while keeping computing costs manageable. We test several machine learning methods—Random Forest (RF), Naïve Bayes (NB), and Classification and Regression Tree (CART). Machine learning refers to algorithms that learn patterns from data to make predictions. Among these, the RF model performs best and is used to assess which variables contribute most to predicting fire occurrence (variable importance analysis). Our goal is to estimate the probability of fire occurrence across Australia and to identify the main drivers of wildfires for the 2019–2020 season. The findings can support more targeted preventive measures in fire-prone areas by focusing on the factors most strongly associated with fire.
[This abstract was generated with the help of AI]
Documents