Flood susceptible prediction through the use of geospatial variables and machine learning methods

Abstract

Floods are one of the most perilous natural calamities that cause property destruction and endanger human life. The spatial patterns of flood susceptibility were assessed in this study using six applied machine learning (ML) models including Decision Tree (DT), Random Forest (RF), Multilayer Perceptron Neural Network (MLP-NN), Adaptive Boosting (AdaBoost), Logistic Regression (LR), and Support Vector Machines (SVM). A flood inventory map was provided using 180 flood points and 180 non-flood points. Both flood and non-flood locations were randomly partitioned into training and testing datasets with the ratio of 70:30. Sixteen various meteorological, hydrological, and geospatial variables were considered in the flood susceptibility assessment. The multi-collinearity analysis was performed to measure the intensity of correlations between the variables and determine whether factors can be included in the flood susceptibility assessment. From a total of sixteen variables, thirteen were chosen based on the multi-collinearity study. The five statistical indices Receiver Operating Characteristics curve (ROC), Jaccard index, F1 score, Overall Accuracy (OA), and Kappa coefficient (K) were used to evaluate models’ performance. Regarding the results of accuracy criteria, the RF model had the best predictive potential. Flood conditioning factors were reclassified using Quantile classification method and the importance of each class from a given flood conditioning factor was evaluated using the Frequency Ratio (FR) method. Additionally, the spatial relationship between the flood conditioning factors and the flood susceptibility map was investigated using the correlation matrix based on the Spearman correlation analysis. Findings of the correlation matrix revealed that the most important flood conditioning factors, in order of importance, were lithology, drainage density, distance from rivers, soil, rainfall, land cover, and Normalized Difference Moisture Index (NDMI). The flood susceptibility map, as the principal output of this study, was produced in five classes ranging from very low to very high susceptibility. The flood susceptibility map provided using the RF model (i.e., the model with the highest accuracy) indicated that the flood susceptibility is very high in 19.73% of the study area, high in 15.49%, moderate in 16.15%, low in 18.72%, and very low in 29.91%. Findings of this research might be an invaluable resource for better management of flooding since it specifically emphasized the primary causes of flooding as well as locations of heightened flood susceptibility.