Hybrid model of spatio-temporal assessment of landslide hazards by using airborne laser altimetry data

Abstract

Landslide is a consistent hazard in mountainous terrain, especially in seismically active areas and regions of high rainfall (tropical regions). The key goal of landslide hazard analysis is to observe the location, intensity (magnitude), and occurrence time (prediction) of landslide. Assessment of landslide hazard and future sensitivity prediction are at the infancy level because of the lack of temporal datasets and efficient modelling techniques. Due to the complex nature of landslides, several modeling methods have been applied for the assessment of landslide hazard. This study developed a hybrid modeling approach named weight of evidence based convolutional neural network (WOECNN) using two models, machine learning model which is convolutional neural networks (CNN) and weight of evidence (WOE) technique of bivariate statistical model. In this study the hybrid model was utilized for generating landslide hazard index (LHI) and sensitivity prediction for future using spatio-temporal data (3D). Two study areas having with environmental and physical conditions including seismo-tectonically active region of Sg. Mesilau, Kundasang, Sabah and Gombak Selangor, were selected for this study. The spatial dataset initially constructed landslide inventory maps for these two sample areas through virtual mapping technique. 21 landslide-causative factors which included slope angle, aspect, altitude, TWI, SPI, TST, TSC, TSConv, soil, lithology, land use, rainfall, seismicity and distance to roads, rivers, roads, and faults were derived from different sources including Airborne Laser Altimetry (ALS) data. These results were validated using receiver operating characteristic (ROC) curve. The area under curve (AUC) of ROC for the landslide hazard index of Sg. Mesilau, Kundasang with respect to different scenarios illustrated that there were variations in between the AUC values for all three scenarios. The AUC value of LHI with maximum parameters was higher (0.82) than the other scenarios as it had less parameters based on priority ranking. While in the Gombak Selangor area the situation was reverse as the AUC values were showed an increasing trend with a decrease in the parameters. The highest AUC value (0.89) was obtained with minimum parameters. Moreover, results of sensitivity prediction (0.86) for Quartz Ridge, Selangor area revealed that the greater part of study was within the range of medium to low hazard. In general, the variation of validation results for landslide hazard assessment in both areas depicted the importance of parameters selection. Based on results, the hybrid models can perform better than the individual models. The present study indicates that WOECNN has performed more efficiently for sensitivity prediction as compared to the individual models.