Machine Learning-Based Wildfire Risk Zoning in the Okavango River Basin: A Transboundary Assessment of Landscape Vulnerability

The Okavango River Basin represents one of the most important transboundary ecosystems in Southern Africa, yet it remains highly vulnerable to wildfires due to the interaction of climatic, topographic, anthropogenic and vegetation-related factors. In support of SDGs 13 (Climate Action) and 15 (Life on Land), this study developed a wildfire risk zoning framework for the basin using a machine learning approach. Multi-source geospatial datasets (2001–2025) were integrated, including burned area products (FireCCI51, 250 m), ignition points from NASA FIRMS, topographic variables derived from SRTM 90 m data, biomass and land use/land cover datasets (500 m) and climatic variables from TerraClimate (~5 km). Data were retrieved in Google Earth Engine. In a QGIS environment with Python integration, pre-zoning layers were generated using a literature-based weighted reclassification scheme, where climatic, vegetation, topographic and historical fire variables were classified into five standardised risk levels according to established thresholds, prior to integration and final classification using the Random Forest algorithm. For the final classification, pre-zoning layers were harmonised to a 90 m spatial resolution in Google Colab. Results revealed the predominance of moderate and high wildfire risk classes, covering more than 570,000 km² of the basin, while approximately 44.77% of the territory was classified as high-risk. Angola (30%) and Botswana (18%) exhibited the largest concentration of fire-prone zones, associated with recurrent fire occurrence, flammable vegetation cover and favourable topographic conditions for fire spread. Conversely, Namibia was characterised predominantly by low to negligible overall fire risk despite elevated climatic susceptibility. Historical fire occurrence presented the highest Mean Susceptibility Index (MSI = 3.62), confirming fire recurrence as the principal conditioning factor of wildfire propagation. The Random Forest model achieved an overall classification accuracy of 72%, demonstrating strong potential for supporting integrated wildfire management, environmental conservation and transboundary fire-risk governance strategies across the study area.