Analysis of Post-flood Impacts on Sentinel-2 Data Using Non-parametric Machine Learning Classifiers: A Case Study from Bihar Floods, Saharsa, India

Abstract

With the advancement of remote sensing and Machine Learning technology, it has become easier to evaluate and map the changes in land cover using high-quality multispectral satellite data. Flood mapping is an important activity for mapping the changes over the given study region for disaster preparedness and carrying out post-disaster mitigation plans. The study’s main objective is to map the waterlogged and silt-affected areas. In this study, three non-parametric Machine Learning classifiers, namely Random Forest (RF), Support Vector Machine (SVM), and k Nearest Neighbors (k-NN), are used for the analysis of post-flood impacts using pixel-based supervised classification technique. Here, Sentinel-2 data (10 m spatial resolution) of the Saharsa district of Bihar, India, have been used for pre-flood and post-flood classification. For the classification process, six Land Use Land Cover (LULC) classes covering a total area of 1667.26 km2 have been identified. All classification results have shown a high Overall Accuracy (OA) ranging from 89% to 92%. The Overall Accuracy and kappa coefficients for pre-flood classified data obtained are RF (89.91%, 0.8772), SVM (89.47%, 0.8718), and k-NN (91.96%, 0.8925). Whereas, for post-flood classified data OA and kappa value obtained are RF (91.54%, 0.8969), SVM (90.94%, 0.8895), and k-NN (89.43%, 0.8712) respectively. Furthermore, quantitative analysis of post-flood classified data revealed a significant increase in waterlogged and silt-affected areas.