The challenge
How to automatically detect African road degradation from satellite imagery, without ground inspection? This challenge was not born in a laboratory. It was born on a road.
Genesis
In 2020, I was involved in a road accident caused by a pothole hidden under mud by villagers. That event planted a conviction: there had to be a way to detect these degradations before they kill. Reducing accident risks, saving lives, this is the founding motivation of TerraPulse Vision.
In 2023, I began formalizing this conviction into a research topic. The hypothesis: a machine learning model trained on satellite imagery can identify potholes and cracks at a scale usable by African authorities. No existing dataset, no reference model for the African context. Just the experience of an accident, a hypothesis, and the determination to validate it.
Scientific approach
The methodology is structured in three phases:
Phase 1, Dataset Building a ground truth dataset from satellite imagery cross-referenced with OpenStreetMap data and direct observations.
Phase 2, Pipeline Developing an image processing pipeline integrating pre-processing, feature extraction and supervised classification.
Phase 3, Model Training and validating a machine learning model for detection and geolocation of degradation zones.
Results
The model demonstrates the technical feasibility of automatic road degradation detection from high-resolution satellite imagery. Supervised classification techniques combined with image segmentation identify degradation zones with operationally viable accuracy. These results constitute the proof of concept that founded Road Intelligence.
What TerraPulse Vision spawned
In 2025, my university's FabLab was looking for a topic on mapping and territorial monitoring to participate in MASS 2025. I proposed TerraPulse Vision. The university validated it and presented it under the name GeoSmart Vision, which won the 2nd Prize at the African Market for Space Solutions. This competitive passage confirmed the viability of the research topic.
The realization came with Road Intelligence: the transformation of this research topic into an operational project, and the first project of the GEOWATCH program. TerraPulse Vision remains the scientific core of Road Intelligence, and a reminder that an idea born from a road accident can become a continental-scale prevention tool.
My contribution
Personal experience at the origin of the project. Research topic formulation. Dataset creation. Processing pipeline design. Machine learning model development and validation. Full authorship of technical specifications. Project proposal for MASS 2025.
Technical resources
Technology badges are displayed below.
- Python
- TensorFlow
- scikit-learn
- Sentinel-2
- PostGIS
- PostgreSQL
- NumPy
- Pandas