Abstract

Wildfires threaten natural ecosystems and human livelihoods in the Tanzanian Miombo woodlands. This study presents the first locally calibrated, high-resolution wildfire susceptibility map for the Iringa region, developed using a robust machine learning framework. Multi-decadal remote sensing data (MODIS fire occurrences, 2001–2024) were integrated with climatic, topographic, vegetation, and anthropogenic variables to train four classifiers: Random Forest, XGBoost, support vector machine with RBF kernel, and Logistic Regression. A balanced dataset of 9096 fire points and an equal number of randomly sampled non-fire points was used. The data were split into 70% for training and 30% for testing. Model performance was evaluated using accuracy, area under the ROC curve (AUC), accuracy, precision, and F1-score. Random Forest achieved the highest overall performance (AUC = 0.845, accuracy = 0.759, precision = 0.789 and F1 = 0.771), followed by XGBoost (AUC = 0.828, accuracy = 0.736, precision = 0.700 and F1 = 0.757), SVM (AUC = 0.755, accuracy = 0.679, precision = 0.648 and F1 = 0.709), and Logistic Regression (AUC = 0.740, accuracy = 0.661, precision = 0.631 and F1 = 0.696). Feature importance analysis identified altitude as the most influential variable, followed by wind speed, distance to road, and NDVI. Kernel Density Estimation revealed spatially distinct fire clusters concentrated in central and southern hotspots. Temporal analysis showed that 94% of fires occur during the dry season (June–November), peaking sharply in October. These findings provide an evidence-based framework for fire prevention and sustainable management of Iringa’s Miombo woodlands.

IPC Classification

Keywords