An integrated machine learning framework for multidimensional cyclone vulnerability assessment and disaster risk reduction in Mozambique

Abstract

Cyclone impacts in data-scarce rural districts are often assessed through fragmented approaches, limiting evidence-based disaster-risk planning. This study developed and externally validated an integrated geospatial framework for Monapo district, Nampula Province, Mozambique, combining an environmental vulnerability index (EVI), a socioeconomic vulnerability index (SVI) at the administrative-post scale, and a Random Forest (RF) classifier. Landsat-8 imagery acquired immediately prior to Cyclone Gombe landfall (January–February 2022) and during the post-landfall period (July–December 2022) was used to derive normalized difference vegetation index, land surface temperature, and a four-class land-cover map, while Shuttle Radar Topography Mission elevation, FAO soils, and 2017 census data were used to construct the EVI and SVI surfaces using fixed 0.2-unit equal-interval classes. The RF model was trained on 1000 stratified samples and internally validated using repeated spatially blocked cross-validation. External validation used the 12 official displacement centres reported by INGC, which were excluded from model fitting. High and very high EVI classes covered 42.3% of the district (2061 km2), whereas high and very high SVI classes covered 28.1% (1370 km2) and included 34.2% of the population (134 724 residents). Model performance was strong, with an area under curve of 0.924 and an overall accuracy of 87.5%. In external validation, 9 of the 12 official displacement centres (75.0%) fell within predicted High or Very High EVI zones, while 8 of the 12 centres (66.7%) were located in High or Very High SVI zones. Post-cyclone settlement expanded by 63 km2, and 67.8% of this growth occurred in moderate-to-high vulnerability zones. The framework provides an evidence base for spatially explicit reconstruction planning and a transferable approach for cyclone vulnerability assessment in data-scarce settings.