Fusion of diverse data sources for flood extent mapping and risk assessment in Sindh: A comparative study of inundation mapping approaches

Abstract

Accurate and timely mapping of flood extent is crucial for effective disaster response and management. However, existing methods often face challenges in integrating diverse data sources and leveraging advanced techniques for precise inundation mapping. This research aimed to address this gap by proposing and evaluating four different methods for improving flood extent mapping in Sindh, Pakistan, through the integration of various data sources, including precipitation, Synthetic Aperture Radar (SAR), buildings, population, and optical data. The first strategy involves categorizing post-flood SAR data immediately using non-parametric (Support Vector Machine (SVM) and Random Forest (RF)), parametric (Maximum Likelihood Estimation (MLE)), and cluster (ISO) algorithms for Vertical transmit and Vertical received (VV) and Vertical transmit and Horizontal received (VH) polarisations. The second method relies on Otsu's method for automatic thresholding on post-flood event VH and VV data. The third approach utilizes both pre- and post-flood event data by creating a stack of Pre-event VH, Pre-event VV, Post-event VH, and Post-event VV data, which is then classified using the SVM, RF, MLE and ISO algorithms. Lastly, the fourth method employs the stack of Principal Component Analysis (PCA) bands, consisting of PCA components 1, 2, and 3, to categorize data for flooding using the SVM, RF, MLE and ISO algorithms. The rationale behind evaluating these diverse methods was to identify the most effective approach for accurately mapping flood extent by exploiting the strengths of different algorithms and data processing techniques. The accuracy and precision of these approaches were rigorously evaluated using Landsat-9 Normalized Difference Water Index (NDWI) as a reference dataset, along with error matrix and compound value (Cv) metrics. The findings show that a PCA-based technique with SVM is marginal superior (based on the overall accuracy and Kappa coefficient with value of 92.20 % and kappa coefficient 0.821) than the other approaches tested in the study. By reducing the dimensionality of the dataset while preserving relevant information, PCA offers a computationally efficient approach to flood extent mapping. Among all the approaches, the collective performance of the algorithms was computed using the Cv metric. The SVM algorithm with a Cv of 1.25 demonstrated the best performance, followed by the RF algorithm with Cv of 1.5, and the MLE algorithm with Cv of 2.5. The worst performing algorithm among all is found to be ISO with Cv of 4. It comes into view that first and second approach underperformed in VH polarisation compared to the VV polarisation and provide a more accurate representation of flooded area in VV polarisation. The flood extent with highest accuracy, together with the world's gridded population, latest Microsoft's Global ML Building Footprints, and OpenStreetMap building data are used in conjunction to estimate the number of people and buildings at risk within the study area.