CABU-Bench : Structured Error Audit of Global Settlement Products across Central African Cities

Sub-Saharan Africa is the world's fastest-urbanizing region, yet Central African cities are the least reliably mapped. Global settlement products (GHSL R2023A GHS-BUILT-S, World Settlement Footprint 2019, Google Open Buildings v3 and 2.5D Temporal, Microsoft Global Building Footprints) are the de-facto labels for mapping, exposure assessment and SDG reporting in this region, but published assessments show their errors concentrate in small, new, peripheral and informal settlement fabric, and our own pre-registration audit measured built-up-area disagreements of ×2.9 to ×3.1 between products in Douala and Yaounde. This study conducts a design-based accuracy assessment and error-structure analysis of these products across twelve Central African cities (Cameroon, DR Congo, Congo, Gabon, CAR, Chad, Equatorial Guinea) at two epochs (T0 = 2018, T1 = 2025), against newly collected two-tier reference data: about 18,000 stratified random point labels (all 12 cities) and dense built-up masks on 60 to 100 tiles in six focal cities. Beyond standard accuracy metrics, the study tests whether product errors are structured: dependent on urban-morphology covariates (H1), spatially autocorrelated (H1), estimable without ground truth from multi-product agreement patterns via latent-class models with modelled cross-product dependence (H2), and correlated across products sharing the same sensor family (H3). Expected outcomes: (i) the first city-level, design-based error atlas of global settlement products for Central Africa; (ii) evidence on the structure of product errors that conditions their safe use as training labels or policy inputs; (iii) public release of the reference benchmark (CABU-Bench v1) with a leakage-controlled spatial evaluation protocol. The study is the first article of a doctoral project on uncertainty-calibrated urban monitoring in cloud-persistent Central African cities from Sentinel-1/2.