Spatio-temporal feature attribution of European summer wildfires with Explainable Artificial Intelligence (XAI)

Hanyu Li; Stenka Vulova; Alby Duarte Rocha; Birgit Kleinschmit

doi:10.1016/j.scitotenv.2024.170330

Spatio-temporal feature attribution of European summer wildfires with Explainable Artificial Intelligence (XAI)

Sci Total Environ. 2024 Mar 15:916:170330. doi: 10.1016/j.scitotenv.2024.170330. Epub 2024 Jan 24.

Authors

Hanyu Li¹, Stenka Vulova², Alby Duarte Rocha³, Birgit Kleinschmit³

Affiliations

¹ Geoinformation in Environmental Planning Lab, Department of Landscape Architecture and Environmental Planning, Technical University of Berlin, 10623 Berlin, Germany. Electronic address: hanyu.li@campus.tu-berlin.de.
² Geoinformation in Environmental Planning Lab, Department of Landscape Architecture and Environmental Planning, Technical University of Berlin, 10623 Berlin, Germany; Institute for Landscape Architecture and Landscape Planning, Department of Environmental Meteorology, University of Kassel, 34127 Kassel, Germany.
³ Geoinformation in Environmental Planning Lab, Department of Landscape Architecture and Environmental Planning, Technical University of Berlin, 10623 Berlin, Germany.

PMID: 38278254
DOI: 10.1016/j.scitotenv.2024.170330

Abstract

Wildfires are among the most destructive natural disasters globally. Understanding the drivers behind wildfires is a crucial aspect of preventing and managing them. Machine learning methods have gained popularity in wildfire modeling in recent years, but their algorithms are usually complex and challenging to interpret. In this study, we employed the SHapley Additive exPlanations (SHAP) value, an Explainable Artificial Intelligence method, to interpret the model and thus generate spatio-temporal feature attributions. Our research focuses on the forest, shrub and herbaceous vegetated areas of Europe during the summers from 2018 to 2022. Using burned areas, meteorology, vegetation, topography, and anthropogenic activity data, we established a wildfire occurrence model using random forest classification. The model was highly accurate, with an Area Under the Receiver Operating Characteristic Curve of 0.940. The SHAP results revealed six features that significantly influence wildfire occurrences: land surface temperature (LST), solar radiation (SR), Temperature Condition Index (TCI), Normalized Difference Vegetation Index (NDVI), precipitation (Prep), and soil moisture (SM). The tipping points for the positive or negative shifts in contributions are around 30 °C (LST), 2.20e⁷ J/m^2 (SR), 0.2 (TCI), 0.78 (NDVI), 2 mm/h (Prep), and 0.18 (SM). These predictors display strong spatial variability in their contribution levels. In Southern Europe, LST and SR emerge as the primary contributors to wildfires, making substantial impacts. Conversely, in regions at mid and high latitudes in Europe, NDVI, Prep, and SM assume a more prominent role in promoting wildfires, albeit with relatively smaller contributions. Furthermore, the disparities in SHAP values for TCI and SMCI across different years provide valuable insights into the effects of variation in regional meteorological conditions on wildfires. Our study provides a new approach to uncovering the spatio-temporal variations of feature contributions, which will help to better understand the mechanism of wildfire occurrence and enhance prevention and mitigation.

Keywords: Driving force; Fire risk; Machine learning; SHapley Additive exPlanation; Spatio-temporal variation; Vegetation.