Automated burned area detection using machine learning with hybrid training data generation and explainable AI: a comparative analysis of PBIA and OBIA approaches

Abstract The combined effects of rising global temperatures and prolonged droughts have led to an increased frequency and severity of wildfires, posing significant threats to ecosystems, economies, and human safety. Rapid and accurate identification of burned areas is therefore critical for effective disaster response, ecological restoration, and resource planning; however, conventional methods are typically reliant on manual labeling or fixed-threshold indices, limiting both their scalability and generalizability. To address these challenges, this study proposes a comprehensive and scalable framework for automated burned area mapping, using Sentinel-2 imagery from the 2022 Marmaris wildfire in Türkiye to evaluate the utility of the approach. The framework developed introduces a hybrid automatic training data generation approach that integrates multiple spectral indices through majority voting with multiscale texture features, eliminating manual labeling and enhancing representativeness. Four tree-based machine learning algorithms were evaluated using both pixel-based image analysis (PBIA) and object-based image analysis (OBIA) schemes, with hyperparameters optimized via Optuna-based NSGA-II multiobjective to achieve a balance between recall and specificity. Shapley Additive exPlanations (SHAP) analysis enhanced transparency in model decisions and identified the most influential features, improving efficiency while maintaining accuracy. The results obtained reveal that OBIA outperformed PBIA with overall accuracies of 98.8% and 93.7%, respectively, and that random forest delivered the most balanced performance. Normalized burn ratio (NBR)-type indices and short-wave infrared (SWIR) bands were identified as the most decisive features, while model-specific dynamic thresholding provided adaptive, data-driven solutions across varying fire conditions. Overall, the proposed framework enhances computational efficiency by reducing processing time and offers a robust and reproducible tool for rapid post-fire damage assessment, ecosystem rehabilitation, and data-driven management in fire-prone regions.