Cyclone surge inundation susceptibility assessment in Bangladesh coast through geospatial techniques and machine learning algorithms: a comparative study between an island and a mangrove protected area

Tropical cyclones, including surge inundation, are a joint event in the coastal regions of Bangladesh. The surge washes out the life and property within a very short period. Besides, in most cases, the area remains flooded for several days. Prediction of inundation susceptibility due to cyclone surge is one of the key issues in reducing cyclone vulnerability. Surge susceptibility can be analyzed effectively through geospatial techniques and various algorithms. Two geospatial techniques, such as GIS-based Analytical Hierarchy Process (AHP) multi-criteria analysis and bivariate Frequency Ratio (FR) techniques, and three algorithms, i.e., Artificial Neural Network (ANN), k -nearest neighbor (KNN) and Random Forest (RF), were applied to understand the comparative surge inundation susceptibility level between an island, i.e., Sandwip and an area protected by mangrove, i.e., Dacope on the Bangladesh coast. A total of ten criteria were considered influential to surge flooding, i.e., Elevation, Slope, Topographic Wetness Index, Drainage density, Distance from river and sea, Wind flow distance, LULC, NDVI, Precipitation, and Soil types. Among them, distance from river and sea (16.34%) and elevation (15.01%) were found to be crucial to surge inundation susceptibility analysis, according to the AHP expert’s opinions. Similarly, precipitation (9.88) and elevation (6.92) in Sandwip and LULC (4.16) and NDVI (4.33) in Dacope were found to be the highest PR values in the FR analysis. The factor maps and final surge susceptibility maps were analyzed through ArcGIS 10.8. The final surge susceptibility maps were categorized into five classes, i.e., very low, low, moderate, high, and very high. Very high and high susceptibility was found around the boundary of Sandwip island and the upper portion of the Dacope upazila. A very high susceptibility area in Sandwip (45.07%) and Dacope (49.41%) was observed by KNN and ANN, respectively. The receiver operating characteristic (ROC) found all techniques acceptable in susceptibility prediction; however, geospatial techniques possessed a better consistent area under the curve (AUC) value than the algorithms for both study sites. Policymakers and professionals can plan to manage disaster reduction activities based on the susceptibility outcomes.