Modeling rules of regional flash flood susceptibility prediction using different machine learning models

Abstract

The prediction performance of several machine learning models for regional flash flood susceptibility is characterized by variability and regionality. Four typical machine learning models, including multilayer perceptron (MLP), logistic regression (LR), support vector machine (SVM), and random forest (RF), are proposed to carry out flash flood susceptibility modeling in order to investigate the modeling rules of different machine learning models in predicting flash flood susceptibility. The original data of 14 environmental factors, such as elevation, slope, aspect, gully density, and highway density, are chosen as input variables for the MLP, LR, SVM, and RF models in order to estimate and map the distribution of the flash flood susceptibility index in Longnan County, Jiangxi Province, China. Finally, the prediction performance of various models and modeling rules is evaluated using the ROC curve and the susceptibility index distribution features. The findings show that: 1) Machine learning models can accurately assess the region’s vulnerability to flash floods. The MLP, LR, SVM, and RF models all predict susceptibility very well. 2) The MLP (AUC=0.973, MV=0.1017, SD=0.2627) model has the best prediction performance for flash flood susceptibility, followed by the SVM (AUC=0.964, MV=0.1090, SD=0.2561) and RF (AUC=0.975, MV=0.2041, SD=0.1943) models, and the LR (AUC=0.882, MV=0.2613, SD=0.2913) model. 3) To a large extent, environmental factors such as elevation, gully density, and population density influence flash flood susceptibility.

Transferability of machine learning-based modeling frameworks across flood events for hindcasting maximum river flood depths in coastal watersheds

“…Intense winds during a hurricane accelerate the movement of floodwaters, leading to greater depths in certain areas, while saturated soil has limited capacity to absorb additional water, resulting in more surface runoff and higher flood depths. The inclusion of elevation as an important feature in our study closely aligns with the findings of Hosseini et al (2020) and Chen et al (2023) in their flash flood susceptibility and hazard assessment one on a small non-coastal watershed and the other on a large coastal watershed.…”

“…We employed common feature selection methods, such as Pearson's correlation coefficients (Cao et al, 2020;Chen et al, 2023;Lee et al, 2020) and principal component analysis (PCA) -a widely used technique in many studies (Abdrabo et al, 2023;Chang et al, 2022;Reckien, 2018) to identify most important features for hindcasting flood depths of a given event in a watershed.…”

“…This may have contributed to Sandy's significant storm surge, which ranged from 1.97 to 2.85 m, compared to Ida and Isaias with storm surge ranges of 0.25 to 0.67 m and 0.20 to 0.76 m, respectively. As a result of Pearson's correlation analyses, we eliminated five features with absolute correlation coefficients greater than 0.70, the cutoff threshold suggested in previous studies (Cao et al 2020;Chen et al 2023;Lee et al 2020). The strong correlation coefficient of 0.99 between "Drainage area" and "Flow accumulation" indicated that both variables capture similar information about water flow and storage in the watershed.…”

“…CC BY 4.0 License.data, we then created a spatially distributed rainfall and wind speed dataset by interpolating the maximum values using the Inverse Distance Weighting (IDW) method(Hosseini et al 2020).Rainfall depth and wind speed are considered for " Point-specific," "spatial average across the contributing watershed," and "spatial maximum across the contributing watershed." These values capture the intensity of the meteorological conditions at individual points and the overall average and maximum values across the watershed.Elevation, ground slope, slope aspect, aspect invariability (ASPVAR), and curvature were features under the topographic category(Cao et al 2020;Chen et al 2023;Huang et al 2022;Khosravi et al 2018;Rafiei-Sardooi et al 2021;Sun et al 2020). DEM with a resolution of 1/3 arc-second (~10 m) was acquired from the United States Geological Survey (USGS 2022).…”