dkemeh1/accra-road-infrastructure-ml

Reproducible ML pipeline for detecting persistent road infrastructure underinvestment in Accra. Accra Road Surface Mapping Pipeline (Sentinel-2 Only) Overview This project provides a fully reproducible workflow for analysing transport infrastructure inequality in Accra, Ghana. The workflow is divided into three integrated parts: - Part 1 — Road Surface Mapping (Python) - Part 2 — Spatial Analysis (QGIS) - Part 3 — Statistical Analysis (Python → Excel) Together, these steps identify blind spots in road infrastructure, assess population exposure, and evaluate transport accessibility inequalities. -------------------------------------------------- PART 1 — ROAD SURFACE MAPPING (PYTHON) What This Part Does This pipeline uses Sentinel-2 imagery and OpenStreetMap (OSM) data to: 1. Extract road surface tags from OSM 2. Generate weak training labels (paved vs unpaved) 3. Segment the road network into 100 m segments 4. Extract Sentinel-2 spectral features 5. Train machine learning models using spatial cross-validation 6. Predict road surface classes 7. Detect road surface changes between years 8. Compare model performance Data Download drive.google.com Place files into: project_root/ ├── snapshots/ ├── imagery/ ├── QGIS-PART 2/ ├── Blindspots_final.csv/ ├── Non_blind_final.csv/ ├── Stable_unpaved_roads_fixed.csv/ ├── Blindspots_districts.csv/ How to Run (Part 1) Step 1 — Install dependencies install_requirements.bat Step 2 — Run pipeline run_pipeline.bat Output (Part 1) outputs/ Includes: - GLOBAL_preprocessing/ - step5_models/ - step6_predictions/ - step7_change_detection/ - _COMPARISON_STEP8/ -------------------------------------------------- PART 2 — QGIS ANALYSIS PACKAGE Overview This section provides the full QGIS environment used to: - Visualise spatial patterns - Identify blind spots - Analyse accessibility - Generate manuscript maps - Export datasets for statistical analysis Contents of QGIS folder: - Accra_QGIS_Project.qgz - GHSL population raster - Blind spot polygons - Stabl …