Nationwide AI Mapping of 43,000+ Buildings Strengthens Disaster Resilience Planning in Saint Vincent and the Grenadines

Small island states like Saint Vincent and the Grenadines face recurring cyclone, flood and landslide risk, but detailed national building-stock data on roof material, pitch and condition is expensive to collect and quickly outdated after each storm.

Under the World Bank's GFDRR Digital Earth Partnership for the Caribbean, researchers aimed to classify roof pitch and material nationwide using satellite imagery and deep learning, to give planners data to target cyclone and flood risk.

Researchers benchmarked geospatial foundation models (Scale-MAE, GASSL with logistic regression/SVM/MLP classifiers) against fine-tuned deep-learning classifiers (ConvNeXt, Vision Transformer variants) on Maxar satellite imagery matched to Microsoft's Building Footprints dataset, training on 3,243 labelled buildings in Saint Vincent and the Grenadines and testing whether pooling in labelled buildings from Saint Lucia and Dominica improved local accuracy.

The fine-tuned ConvNeXt-large model trained on the pooled regional dataset reached an F1 score of 0.878 for roof-pitch classification, and a locally trained model reached F1 0.835 for roof-material classification; deep-learning models outperformed pretrained foundation-model baselines. The winning model was applied nationwide, producing classified building-footprint maps for all 43,061 structures in the country.

Published in September 2025 as a technical research output; no public source yet confirms the classifications have been formally adopted into a government building code, insurance scheme or planning-decision process.