Gamuchirai-Magamba/zw-vulnerability-map

Ward-level child marriage estimates for Zimbabwe, benchmarked to survey district estimates # zw-vulnerability-map **Zimbabwe's child marriage estimates, from 91 districts down to 1,961 wards.** Satellite imagery, a geospatial foundation model, and estimates that reconcile exactly with the survey data they came from. ### 🗺️ Open the map · API docs · Model card --- ![The map]screenshot.png) *Chipinge: 30 wards spanning 36–70%, against a district estimate of 55.8%. The four ⚠️ wards fall outside anything the model was trained on. The API sleeps after 15 minutes of no traffic, so the first request may take up to a minute.* --- ## The problem this exists for `zw-marriage-risk` answered *which districts*. It found that one in three Zimbabwean girls marries before 18, and estimated a rate for each of the 91 districts. **A district is not where a programme happens.** Mbire holds roughly 100,000 people spread over an area larger than some countries. An organisation with funding for twelve wards still has to guess which twelve. Zimbabwe has **1,961 wards**, and we cannot survey our way to them. The 2019 MICS interviewed 10,703 women across 462 clusters — spread over 1,961 wards that is five women per ward, and most wards have none. ## The idea **If we cannot measure every ward, we can learn what a vulnerable place looks like and then look at every ward from space.** The DHS and MICS surveys give GPS coordinates for **862 clusters** — real villages and neighbourhoods with real interviewed women. That is the training set. Each one is described by four contextual layers and by Google's Satellite Embedding, a model trained on satellite imagery itself. ## The rule that keeps it honest > **The satellite model does not decide how bad a district is. The survey model > decides that. The satellite model only decides which parts of a district are > worse than others.** Every district's wards are rescaled until their population-weighted mean equals the district estimate from Phase 2: ``` factor_d = phase2_estimate_d / pop_weighted_mean(raw wards in d) ad …