Page 17 - TU Dresden - StuFoExpo 2024: Book of Abstracts

Enhanced Wildfire Susceptibility Mapping by Integrating Remote
Sensing, Reanalysis Data, and Machine Learning Approaches
Case study: Saxony State, Germany

Osmani, Hiva | hiva.osmani@mailbox.tu-dresden.de
Fakultät Umweltwissenschaften, Technische Universität Dresden
_____________________________________________________________________________________________________

To address the global concern about wildfires and their socioecological consequences, cost-effec-
tive and practical technologies to reliably predict wildfire outbreaks are urgently needed. The study
assessed the effectiveness of Random Forest and Generalized Linear Model in mapping complex
wildfire-prone terrain in Saxony, Germany, and identified key factors influencing mapping. To this
end, the Random Forest (RF) and Generalized Linear Model (GLM) algorithms applied to predict the
wildfire susceptibility map.

Overall, 315 historical wildfires were identified from 2010 to 2020 in the Saxony state, Germany.
Among them, 221 events (70%) were randomly selected to generate the models, and the remaining
94 (30%) were used for validation. 10 influencing factors: elevation, aspect, slope, landcover, topo-
graphic wetness index, mean annual wind speed, mean annual air temperature, mean monthly
rainfall, distance to the river, and normalized difference vegetation index were selected to generate
the proposed models and detect fire-prone areas. Area Under the Curve (AUC) of the Receiver
Operating Characteristic (ROC) was computed to verify the reliability and accuracy of the wildfire
susceptibility maps. With the GLM attaining an AUC-ROC of 0.799 and an accuracy of 71.1%, and
the Random Forest model closely trailing with an AUC-ROC of 0.791 and an accuracy of 70.6%, both
models showed robust ability to distinguish between wildfire and non-wildfire regions.

The study used the Random Forest model to assess the impact of environmental and geographic BAU UND UMWELT
factors on wildfire susceptibility, finding temperature as the most critical factor, followed by rain-
fall, elevation, and landcover. Overall, the proposed frameworks can be used to anticipate the spa-
tiotemporal patterns of wildfire occurrence, which is crucial for land management, preventing wild-
fires, and reducing their effects.

The flowchart of developing the fire susceptibility map using ML approaches.

12