yuzukumo/ssa-urban-deprivation-benchmark

Open multimodal benchmark for measuring and modeling intra-urban deprivation across sub-Saharan African cities using public spatial data, weak supervision, and external validation. SSA Urban Deprivation Benchmark Transfer evaluation of public-data deprivation proxies across sub-Saharan African cities. ## Research Question When a target city has no local survey labels, can public spatial data provide a useful and transferable deprivation proxy? The project compares an interpretable infrastructure-scarcity atlas with spatial interpolation, gradient boosting, convolutional models, multimodal fusion, and a spatial graph model. The central comparison is deliberately conservative: a complex model is useful only when it improves on both the atlas and a coordinates-only spatial interpolation baseline under a transfer protocol that prevents local spatial leakage. ## Study Design | Component | Primary benchmark | Transfer extension | | --- | --- | --- | | Cities | 7 cities | 14 cities in 7 countries | | Grid | 500 m cells | 500 m cells | | Benchmark support | 15,030 rows in GHS-SMOD classes 21, 22, 23, and 30 | 23,273 rows in GHS-SMOD classes 21, 22, 23, and 30 | | Main protocol | 8-seed leave-one-city-out | Leave-one-city-out, leave-one-country-out, and same-country transfer | | Evaluation unit | Primary RWI support cluster | Primary RWI support cluster | The core7 cities are Accra, Addis Ababa, Dar es Salaam, eThekwini, Kampala, Nairobi, and Port Harcourt. The core14 extension adds Dire Dawa, Enugu North, Jinja, Johannesburg, Mombasa, Mwanza-Nyamagana, and Tamale. The weak target is the negative raw Relative Wealth Index (RWI). RWI is relative within country and is not an absolute cross-country wealth scale. Regression targets are standardized within each training fold; classification is defined as the lowest 20% of reliable RWI values within each city. Metrics are aggregated by primary RWI support rather than treating interpolated 500 m cells as independent survey observations. ## Results ### Core7: Eight-seed LOCO Values are unweighted means of city-level support-cluster metrics across eight seeds. RMSE is lower-is-better; Sp …