Prediction of landslide susceptibility using Hybrid Support Vector Regression (SVR) with metaheuristic algorithms

Abstract

<p>Spatial landslide susceptibility prediction is essential for adopting landslide mitigation strategies and reducing landslide damages. This study proposes novel hybrid models based on support vector regression (SVR) and meta-heuristic algorithms: Gray Wolf Optimization Algorithm (GWO) and Artificial Bee Colony Algorithm (Bee). Geospatial data including 15 environmental landslide conditioning factors (slope, aspect, plan curvature, air flow, convergence index, terrain surface texture, wind exposition index, vector ruggedness measure, TWI, valley depth, forest type, forest density, forest age, geology and land use) were derived for a landslide-prone region of Icheon, Korea from available data. A landslide inventory map with 457 landslide points was created from existing aerial photos and field surveys. The geospatial data and landslide points were divided to training (50%) and validation (50%) dataset and used to construct the landslide susceptibility models using the SVR model. The parameters of the SVR models were optimized using the GWO and Bee algorithms. The resultant hybrid models (SVR-GWO and SVR-Bee) leveraged the advantages of the GWO and Bee meta-heuristic algorithms for parameter estimation. The predictive accuracy of the models was quantified using the statistical measures of RMSE, MAE, AUC, and ROC curve. Both GWO and Bee algorithms improved the SVR performance with SVR-GWO model performing the best (AUC = 0.82) followed by SVR-Bee model (AUC = 0.82) and standalone SVR model (AUC = 0.79). The results demonstrated the efficiency and improved performance of the proposed hybrid models compared to standalone models for spatial landslide susceptibility prediction with limited environmental data.</p>