Dodokal/telecom-tower-siting-Ethiopia

# Machine Learning and Geospatial Modeling Reveal Telecommunication Tower Suitability, Deployment Priorities, and Digital Connectivity Gaps Across Ethiopia This repository contains the code, predictor metadata, and reproduction instructions for the manuscript: > **Reja, A. A. & Yasin, K. H. (2026).** *Machine Learning and Geospatial Modeling Reveal Telecommunication Tower Suitability, Deployment Priorities, and Digital Connectivity Gaps Across Ethiopia.* Manuscript under review. The framework integrates open Earth observation, Demographic and Health Survey (DHS), and crowdsourced infrastructure data through a stacked ensemble of LightGBM, XGBoost, Random Forest, and MaxEnt classifiers under nested spatial block cross-validation. It produces a national 1 km suitability surface, an operational priority surface, and a digital-equity-adjusted priority surface that systematically reveals geographies of digital exclusion missed by conventional siting maps. --- ## Quick links - 📄 Manuscript and supplementary materials - 🔬 Reproduction guide - 🛠️ Pipeline scripts - 📊 Headline results - 📚 Data sources - ✉️ Contact --- ## Headline results | Metric | Value | |---|---| | Tree-based AUC range (10-fold spatial CV) | **0.963–0.970** | | LightGBM Brier score | **0.001529** | | Greenfield-priority area | **109,630 km² (9.87% of national)** | | Densification-priority area | **6,054 km² (0.54%)** | | Spearman ρ (priority vs equity gap) | **−0.037** (essentially orthogonal) | | Regional rank stability across 4 classifiers | **ρ = 0.91** | | Variogram range of residuals | **7.45 km** (≪ 50 km block size) | See `results/` for the full set of cross-validated metrics, Moran I diagnostics, calibration scores, and threshold-sensitivity tables. --- ## Repository structure ``` telecom-tower-siting-ethiopia/ ├── README.md ← you are here ├── LICENSE ← CC BY 4.0 ├── CITATION.cff ← machine-readable citation metadata …