Enhancing wildfire mapping accuracy using mono-temporal Sentinel-2 data: A novel approach through qualitative and quantitative feature selection with explainable AI

Accurate wildfire severity mapping (WSM) is crucial in environmental damage assessment and recovery strategies. Machine learning (ML) and remote sensing technologies are extensively integrated and employed as powerful tools for WSM. However, the intricate nature of ML algorithms often leads to ‘black box’ systems, obscuring the decision-making process and significantly limiting stakeholders' ability to comprehend the basis of predictions. This opacity hinders efforts to enhance performance and risks exacerbating overfitting. This present study proposes an innovative WSM approach that incorporates qualitative and quantitative feature selection techniques within the Explainable AI (XAI) framework. The methodology aims to enhance the precision of WSM and provide insights into the factors contributing to model decisions, thereby increasing the interpretability of predictions and streamlining models to improve performance. To achieve this objective, we employed the SHapley Additive exPlanations (SHAP)-Forward Stepwise Selection (FSS) method to demonstrate its efficacy in elucidating the qualitative and quantitative impacts of predictors on ML algorithm performance, accuracy, and interpretability designed for WSM. Utilizing post-fire imagery from Sentinel-2 (S2), we analyzed ten bands to generate 225 unique spectral indices utilizing five different calculations: normalized, algebraic sum, difference, ratio, and product forms. Combined with the original S2 bands, this resulted in 235 potential predictors for ML classifications. A random forest model was subsequently developed using these predictors and optimized through extensive hyperparameter tuning, achieving an overall accuracy (OA) of 0.917 and a Kappa statistic of 0.896. The most influential predictors were identified using SHAP values, with an FSS process narrowing them down to the 12 most critical for effective WSM, as evidenced by stabilized OA and Kappa values (0.904 and 0.881, respectively). Further validation using a ninefold spatial cross-validation technique demonstrated the method's consistent performance across different data partitions, with OA values ranging from 0.705 to 0.894 and Kappa values from 0.607 to 0.867. By providing a more accurate and comprehensible XAI-based method for WSM, this research contributes to the broader field of environmental monitoring and disaster response, underscoring the potential of integrated qualitative and quantitative analysis to enhance ML models' capabilities.