A Satellite-Based Framework for Flood Mapping: Enhancing Explainability Through Multispectral Indices and SHAP Analysis

While machine learning (ML) models have shown considerable promise for modelling complex environmental phenomena such as floods, their adoption in flood mapping is often constrained by limited interpretability. This study develops an integrated SHapley Additive exPlanations (SHAP)-enhanced framework for satellite-based flood extent mapping using multispectral and topographic variables. Spectral information derived from Landsat 7, including the Normalized Difference Water Index (NDWI), Automated Water Extraction Index (AWEI), Normalized Difference Vegetation Index (NDVI), and near- and shortwave-infrared reflectance, was combined with slope and elevation derived from SRTM. Historical flood extent labels were obtained from the Global Flood Database. Feature selection was performed in two stages by first reducing redundancy through correlation analysis and then applying SHAP-based attribution to quantify the contribution of the remaining variables. The final model retained four primary predictors: AWEI, Near-Infrared (NIR) reflectance, slope, and elevation. A supervised Light Gradient Boosting Machine (LightGBM) model achieved an accuracy of 92% and an AUC of 0.950, outperforming Logistic Regression and TabNet. SHAP analysis provided global and local explanations of the model output, while spatial comparison with the MODIS-derived flood extent showed broad agreement in the analysed event. The framework therefore focuses on event-based flood extent classification rather than long-term pre-event flood susceptibility prediction and demonstrates the value of combining predictive performance with model transparency for remote-sensing-based flood mapping.