Landslide spatial modelling using unsupervised factor optimisation and regularised greedy forests

This study evaluates the contribution of an unsupervised factor optimisation based on sparse autoencoders (SAEs) to spatial landslide modelling with regularised greedy forests (RGFs). A total of 952 landslides were identified by field surveys, equally divided and used for training and testing of the proposed model. Ten conditioning factors related to landslides, including geo-morphometrical (i.e. altitude, slope, aspect, curvature, slope length, topographic wetness index and sediment transport index) and geo-environmental (i.e. lithology, nearness to roads and nearness to streams), were used to investigate the spatial relationships between the variables and landslides. 'The steps of the modelling were twofold. First, the factors were optimised by SAE to reduce information redundancy and correlation in the data. Second, RGF was used to create landslide susceptibility maps with the optimised feature representations. The area under the receiver operating characteristic curve (AUROC) was used to assess the predictive ability of the proposed models. Experimental results show that the proposed SAE-RGF outperforms the RGF and random forest (RF) models in terms of prediction rate and is less sensitive to overfitting and underfitting. The highest prediction rate (AUROC = 0.892) was obtained with only seven features by the SAE-RGF model, which is better than the two other methods (RGF and RF). The unsupervised factor optimisation approach not only reduces computation time but also improves the prediction accuracy of tree based models, including RGF. The generated landslide susceptibility maps can be implemented to mitigate landslide hazards and to designate land use by stakeholders (e.g. planners and engineers).