A new forecasting approach using the combination of machine learning to predict flood susceptibility (case study: Karun catchment)

Due to its environmental diversity, Iran ranks high in terms of crises caused by natural disasters. Flooding, as one of these disasters, is  causing severe social, economic, health, and environmental damage in many areas due to rapid urban growth and climate change. Therefore spatial forecasting of floods is crucial, as failure to identify flood risk areas in a catchment can exacerbate the destructive effects of floods. Recent advances in remote sensing, geographic information systems, machine learning, and statistical modelling have made it possible to produce highly accurate flood prediction maps. This study aims to predict flood risk areas in the Karun watershed using Sentinel satellite images and a novel ensemble approach with six machine learning models.

In this study, Synthetic Aperture Radar (SAR) data from Sentinel-1 images were used to identify areas affected by flooding.  First, the dates of heavy rainfall and flooding events in the study area were identified from various sources of information. Subsequently, Sentinel-1 images were obtained from the Copernicus database, representing the area before and after the flood events. The aforementioned data were processed using the SNAP platform. The identification of flood-affected areas was achieved through the application of the thresholding technique. For this purpose, the Normalized Difference Water Index (NDWI) generated from Sentinel-2 images and land cover classes indicating permanent water bodies were employed to determine the threshold for identifying flood-affected areas. The flood polygon layer was converted to a point layer, resulting in a total of 70 flood occurrence points. A review of previous studies and local characteristics identified seven main factors that significantly affect flood occurrence in the region. These factors include the Normalized Difference Vegetation Index (NDVI), Topographic Wetness Index (TWI), slope, flow direction, flow accumulation, distance from the river, and monthly rainfall. Additionally, the Digital Elevation Model (DEM) of the region was obtained from the SRTM database, and the spatial resolution of all factors was aligned with the DEM layer. Subsequently, various machine learning algorithms were employed to develop a combined model that provides more accurate predictions of flood-prone areas. The individual models include the Generalized Linear Model (GLM), Boosted Regression Tree (BRT), Support Vector Machine (SVM), Random Forest (RF), Multivariate Adaptive Regression Splines (MARS), and Maximum Entropy (MAXENT).                    

The results of this study indicate that the northeast of Aligudarz city, parts of Durud and Azna in Lorestan province, Khademmirza, Shahrekord, and Kiyar in Chaharmahal Bakhtiari province, Dana and Boyer Ahmad in Kohgiluyeh and Boyer Ahmad province, Semirom city in Isfahan province, and the southern border areas of Karun River in Khuzestan province have the highest flood potential in this basin. The performance evaluation of the models revealed that the Random Forest (RF) and Maximum Entropy (MaxEnt) models exhibited the highest accuracy among the individual models. These models, by combining environmental information and flood occurrence data, can produce highly accurate flood susceptibility maps. These maps can serve as crucial management tools to mitigate the adverse effects of floods and prevent development in vulnerable areas.

Overall, this study demonstrates that the use of an ensemble approach which combines machine learning models can provide more reliable results in the prediction of flood risk areas. The findings of this research are beneficial for managers and planners, as they can prevent development in vulnerable areas and consequently help reduce financial losses and human damages in the future.