Rural Household Landslide Loss Estimation in Peru: An Exploratory Neural-Network Approach

Abstract

This work-in-progress paper proposes an exploratory Neural-Network framework to estimate household-level economic losses from debris-flow-induced landslides in two rural communities of Carabayllo, Lima, Peru. The study integrates geophysical, geographical, environmental, and sociodemographic information from georeferenced household surveys and geomorphological outputs generated with the LAPSUS model (v6.03). The main challenge is a small, heterogeneous, and partially interdependent dataset (N = 60 households) comprising both discrete and continuous variables. To address this complexity, the framework combines autoencoder-based dimensionality reduction for discrete variables, transformed feature-space construction, unsupervised segmentation, and neural-network regression. Rather than claiming validated predictive performance, the paper presents this architecture as a proof of concept for structurally complex disaster datasets under severe sample constraints. It contributes methodologically by showing how representation learning and segmentation may support loss estimation, and empirically by highlighting the close relationship between social and housing conditions and household disaster losses in vulnerable peri-urban settings.