Neighbourhood-Scale Typhoon Risk Mapping Using Machine Learning and SAR-Derived Impact Proxies: Cross-Event Evidence from Cebu City, Philippines

Abstract

Typhoon impacts in the Philippines exhibit strong neighbourhood-scale heterogeneity driven by interactions among hazard intensity, exposure, and vulnerability, yet most operational assessments remain hazard-centric. This study presents a grid-based machine-learning framework for neighbourhood-scale typhoon risk mapping integrating geophysical hazards, exposure and vulnerability indicators, and satellite-derived impact proxies. A Random Forest classification model is implemented at 100 m resolution for Cebu City, using Sentinel-1 Synthetic Aperture Radar (SAR) backscatter change as an all-weather indicator of disturbance. Binary impact labels are generated by applying a disturbance threshold (τ = 1.5) to SAR backscatter change, converting the disturbance signal into a super-vised classification target. Because SAR backscatter changes may reflect surface wetness, vegetation disturbance, or flooding effects, the label represents a disturbance proxy rather than a direct measure of structural damage. Predictor variables include ERA5-Land wind and rainfall metrics, population and building exposures, topography, land cover, and proximity to coastal and riverine features. Model performance is evaluated using the Area Under the Receiver Operating Characteristic Curve (AUC) through cross-event generalisation between Typhoon Rai (2021) and Typhoon Phanfone (2019). Results show above-random discrimination (average AUC ≈ 0.59) and stronger transferability in built-up areas (AUC ≈ 0.69), supporting neighbourhood-scale risk screening under realistic data constraints.