An evolutionary approach for spatial prediction of landslide susceptibility using LiDAR and symbolic classification with genetic programming

This research examines the potential of spatial prediction of landslide susceptibility by implementing an evolutionary approach using symbolic classification with genetic programming (GP). Specifically, the light detection and ranging (LiDAR)-based digital elevation model was used to generate topographic prediction attributes and to digitize the location of shallow landslides by derivatives such as hillshade maps and contours. The presented approach tested a total of 72 runs with different parameter configurations for producing a good outcome among a number of possible solutions by varying population size, tournament group size and mutation probability. The final solution depicted a total of three important variables including slope, wetness index and solar insulation that were used in the prediction. The GP methodology used symbolic expression trees for the development of the predictive models that were tested and validated in the northern portion of the Cuyahoga Valley National Park located in northeast Ohio. The selected solution from the implemented approach showed that the area under the curve from the receiver operating characteristic curves had a high discrimination power in separating the areas with high susceptibility. The presented model yielded an accuracy of 85.0 % classifying a total of 13.4 % as high susceptibility area with an overall quantitative index of accuracy corresponding to 0.9082. Based on obtained results, the potential of the presented GP approach for mapping landslide susceptibility is promising and further exploration of its capabilities is suggested for finding new avenues of possible landslide research and practical implementations.