ARTIFICIAL NEURAL NETWORKS FOR LANDSLIDE SUSCEPTIBILITY MAPPING: SCALABLE APPLICATIONS FROM PERU TO MINING OPERATIONS *A. Otiniano1,2, J. Andrade2, C.N. Fernández3, J.A. Rosales2 1Digital Innovation Unit, Earth & Environment Division, WSP Peru, Lima, Peru (*Presenting author: alonso.otiniano@wsp.com ) 2Multidisciplinary Sensing, Universal Accessibility and Machine Learning Group, National University of Engineering, Lima 1533, Peru 3Directorate of Mineral and Energy Resources, Systematization of Economic Geological Information, INGEMMET, Lima, Peru ABSTRACT Peru's diverse geography and topography make it susceptible to frequent and devastating landslides. Although numerous studies have mapped landslides across the country, existing susceptibility maps lack quantitative methods, limiting their effectiveness. This research addresses that gap by developing a new quantitative map of the susceptibility to landslides for the Peruvian territory using a novel approach based on Artificial Neural Networks (ANN). ANN were selected for their ability to capture complex, non-linear relationships between various geomorphometric variables. The model was trained using the resilient backpropagation algorithm with weight backtracking and topographic and geomorphological factors as input data, including the digital elevation model, land cover (7 classes), slope, aspect (8 classes), curvature, topographic wetness index (TWI) and stream power index (SPI). A total of 105,181 data points representing landslide and non-landslide locations were used in the modeling process. The ANN model achieved a training area Under the Curve (AUC) of 97.37% and a testing AUC of 95.55%, demonstrating its high performance. Furthermore, when applied to new data, the ANN-derived landslide map identified that 73.74% of the total data points in the study area fall within the "very high" or "high" susceptibility, significantly exceeding the 37.24% and 48.37% identified by the Macro-regional and Regional models of Peru developed in 2010 and 2018, respectively. A preliminary down-scaling experiment was conducted using a high-resolution dronederived DEM (0.25 m × 0.25 m) from a confidential pilot site. Results indicate that the model preserves predictive consistency when transferred to local scale and is applicable to mine-scale hazard assessments while enhancing morphometric detail. This behavior is consistent with the selfaffine nature of topographic surfaces, whose statistical structure remains scale-invariant. By approximating a continuous non-linear relationship between terrain attributes and landslide probability, the ANN framework benefits directly from improved topographic fidelity without requiring architectural modifications or retraining. Comprehensive down-scaling validation is
RkJQdWJsaXNoZXIy MTM0Mzk2