Assessing Machine Learning Models on Temporal and Multi‐Sensor Data for Mapping Flooded Areas

Natural disasters, particularly floods, are escalating in frequency and intensity, disproportionately impacting economically disadvantaged populations and leading to substantial economic losses. This study leverages temporal and multi‐sensor data from Synthetic Aperture Radar (SAR) and multispectral sensors on Sentinel satellites to evaluate a range of supervised and semi‐supervised machine learning (ML) models. These models, combined with feature extraction and selection techniques, effectively process large datasets to map flood‐affected areas. Case studies in Brazil and Mozambique demonstrate the efficacy of the methods. The Support Vector Machine (SVM) with an RBF kernel, despite achieving high kappa values, tended to overestimate flood extents. In contrast, the Classification and Regression Trees (CART) and Cluster Labeling (CL) methods exhibited superior performance both qualitatively and quantitatively. The Gaussian Mixture Model (GMM), however, showed high sensitivity to input data and was the least effective among the methods tested. This analysis highlights the critical need for careful selection of ML models and preprocessing techniques in flood mapping, facilitating rapid, data‐driven decision‐making processes.