Flood susceptibility mapping using machine learning and remote sensing data in the Southern Karun Basin, Iran

Floods in Iran lead to significant human and financial losses annually. Identifying flood-prone regions is imperative to minimize these damages. This study aims to pinpoint flood-susceptible areas in the Great Karun Plain using remote sensing data, Google Earth Engine (GEE), and machine learning techniques. For the analysis, Landsat 8 data from April 8, 2019, and multiple variables including actual evapotranspiration, aspect, soil bulk density, clay content, climate water deficit, elevation, NDVI, land cover, Palmer Drought Severity Index, reference evapotranspiration, precipitation accumulation, sand content, soil moisture, minimum temperature, and maximum temperature were employed. These variables were utilized in the machine learning process to establish flood susceptibility zones. During the machine learning process, the base flow data of the Karun River was extracted from the Landsat image. A total of 19,335 samples were incorporated into the machine learning procedure using techniques such as MARS, CART, TreeNet, and RF. The model assessment criteria encompassed ROC, sensitivity, specificity, overall accuracy, F₁score and mean sensitivity. Results indicated that the TreeNet technique yielded the most promising outcomes among the machine learning algorithms with ROC values of 0.965 for test data. The characteristic criterion reached 91.2%, while the overall accuracy criterion stood at 91.12%. The model’s average sensitivity was 90.81%, and F1score was 63.51% for this technique. Additionally, analysis of the relative importance of independent variables highlighted that factors like vegetation cover (0.37), cumulative precipitation (0.23), soil water deficit (0.12), drought intensity (0.12), and landscapes (0.1) exerted a more pronounced influence on flooded areas compared to other variables.