Addressing data imbalance in urban informal settlement mapping from earth observation using ensemble learning: A case study in Rio de Janeiro

Informal settlements pose major challenges for public health, infrastructure, and urban planning due to their high density and unregulated growth. Remote sensing has emerged as a key tool for mapping these areas, but strong class imbalance -where informal settlements represent a small fraction of urban land -remains a critical barrier. Existing methods often rely on simple undersampling, discarding valuable training data from formal residential zones. We propose BALISE, a novel ensemble learning approach that leverages the full extent of available data to improve informal settlement detection from remote sensing. Our framework combines Sentinel-2 multispectral imagery and the Copernicus Digital Elevation Model with auxiliary features derived from OpenStreetMap. We evaluate our method on a use case in Rio de Janeiro, using a spatial crossvalidation strategy that rigorously tests generalization across five different urban zones. BALISE improves both F1-score and Kappa coefficient by approximately 2 points over standard undersampling, offering a robust and transferable tool for remote sensing-based urban analysis in fragmented and socially complex environments.