<![CDATA[This research introduces a framework for classifying flood inundation utilising Sentinel-1 Ground Range Detected (GRD) radar imagery alongside machine learning algorithms. Radar backscatter values from pre- and post-event Sentinel-1 images were processed with SNAP and QGIS to extract spatial features and change indicators in decibel (dB) format. The tabular dataset, comprising 500,000 samples that equally represent flooded and non-flooded areas, was utilised for model training. Two models, Random Forest and Naive Bayes, were assessed for their classification efficacy. The Random Forest model demonstrated exceptional performance, attaining an accuracy of 99.81%, precision of 99.75%, recall of 99.67%, and an F1-score of 99.71%. Naive Bayes achieved an accuracy of 52.63%, with precision and F1-score notably impacted by elevated false positive rates, although recall was 86.36%. Analysis of confidence distribution indicated that Random Forest exhibited low-confidence errors at the decision boundary, whereas Naive Bayes demonstrated confident misclassifications. Analysis of computation time indicated that Naive Bayes required less than 0.1 seconds per run, whereas Random Forest completed training in under 3 minutes. The trade-off between speed and reliability underscores the appropriateness of Random Forest for operational flood mapping applications. This research provides a practical comparison of classification models utilising open-access radar data and establishes a dependable pipeline for pixel-level flood identification.]]>
