zrmondsc/uganda_roads

automated extraction of roads and footpaths in nakivale refugee settlement in Uganda...and more?? # **uganda_roads** ### *Automated road and footpath extraction in Ugandan refugee settlements using VHR multispectral imagery and lightweight CNNs.* This repository contains the data processing pipeline and preprocessing scripts for a deep-learning workflow designed to extract road and footpath networks from **very high-resolution (0.5 m) Maxar imagery** in **Nakivale Refugee Settlement**, southwestern Uganda. My research questions are threefold: 1. Can a lightweight CNN architecture, originally developed and benchmarked on urban road systems, be adapted to humanitarian settings where infrastructure is informal and sometimes ephemeral? 2. Can we train a model using annotations for roads and footpaths that are adapted from OpenStreetMap data, despite incomplete and uneven coverage in the training dataset? 3. Does adding small amounts of high-quality annotations significantly improve model accuracy? Although model training has not yet begun, this repository includes: - An **HPC-compatible preprocessing pipeline** - Automated **Maxar scene indexing**, clipping, cutline generation, and tiling - Integration with **OpenStreetMap (OSM)** road data via `osmnx` - A GitHub compatible repository structure designed for future modeling tasks --- ## **Project Motivation** Refugee settlements are among the fastest-growing and least-mapped landscapes in the world. In Uganda, more than **1.4 million refugees** live in refugee settlements, yet maps of infrastructure are often incomplete or outdated in OpenStreetMap and global datasets. Accurate maps of roads and footpaths are important for: - Modeling mobility and accessibility - Understanding settlement morphology - Identifying other data inequities The performance of deep learning road extraction in **rural humanitarian contexts** is still largely unexplored. Settlements like Nakivale contain unpaved, highly diverse, and sometimes ambiguous infrastructure networks. This project investigates whether modern CNNs can address …