nicoflightsystems/terrain-slope-open-buildings

Reproducible analysis of terrain slope and Google Open Buildings detection recall in highland Ethiopia. # Terrain slope and open building-footprint completeness Reproducibility materials for a study comparing Google Open Buildings v3 with hand-labelled ORBITaL-Net reference buildings in highland Ethiopia. The analysis asks whether building-detection recall varies with terrain slope after controlling for tree cover and reference-imagery acquisition year. The study measures an association; it does not identify a single optical or machine-learning mechanism. ## Reproduced headline result The checked-in processed tables reproduce the paper's primary sample and model: - 1,269 chips with at least one reference building and verified detection coverage - 16,336 reference buildings after the pre-specified 10-500 m2 area filter - 12,770 matched detections - sample recall: 78.17% - terrain-slope effect: approximately -7.5 percentage points per +10 degrees in the pooled model - terrain-slope effect: approximately -5.7 percentage points per +10 degrees with source-scene fixed effects These are model-implied discrete probability changes evaluated at the sample mean recall, not average marginal effects. Confidence intervals use source-scene clustered sandwich covariance. ## Quick reproduction ```bash python3 -m venv .venv source .venv/bin/activate pip install -r requirements.txt python src/reproduce_results.py python src/verify_reproduction.py ``` Generated tables and figures are written to `results/`. The same reproduction check runs automatically on GitHub Actions after every push. Dependencies for the recovered geospatial-preparation scripts are listed separately in `requirements-geospatial.txt`. ## Repository layout ```text data/processed/ Small derived tables needed to reproduce the models docs/DATA_SOURCES.md Upstream datasets, licenses, and acknowledgements src/reproduce_results.py Exact grouped-binomial GLMs and figures src/verify_reproduction.py Regression checks against the manuscript values src/recovered_pipeline/ Original data-preparation scripts …