Utilizing Satellite Imagery Datasets and Machine Learning Data Models to Evaluate Infrastructure Change in Undeveloped Regions

Kyle McCullough, Andrew Feng, Meida Chen, Ryan McAlindenarXiv 2020 · arXiv

Abstract

Monitoring infrastructure development in remote regions is often hindered by the sheer volume and cost of high-resolution satellite imagery. This paper proposes a predictive framework that combines game-engine navigation meshes (NavMeshes) for high-precision local routing and Deep Inverse Reinforcement Learning (IRL) for long-range trajectory estimation. By modeling engineering constraints like track grade and terrain geometry, the system identifies the most likely paths for future construction, allowing analysts to target imagery requests more efficiently. The framework was validated using the Addis Ababa-Djibouti Railway, showing that low-resolution elevation data can effectively guide high-resolution monitoring efforts.

Introduction: The "Data Needle in a Haystack"

The rapid expansion of infrastructure in developing nations presents a significant monitoring challenge. While satellite imagery archives now contain hundreds of petabytes of data, manually identifying progress on projects like transcontinental railways is prohibitively expensive and time-consuming. High-resolution imagery (30-50cm GSD) is necessary for detailed analysis but carries high costs and storage requirements.

The core issue is predictive: analysts often do not know where to look next. Without a model to predict the most likely path of a railway under construction, vast areas must be imaged and searched. This research addresses this gap by automating the monitoring process through terrain-based path prediction.

The study area in Ethiopia and Djibouti, showing the satellite data request region for the railway project.

Methodology

The authors developed a dual-approach framework to handle different scales of infrastructure prediction.

1. Navigational Mesh (NavMesh) Pathfinding

For short-distance, high-precision routing, the team utilized tools from the commercial game industry. By ingesting 3D satellite-derived meshes into the Unity game engine, they applied the A* Pathfinding algorithm. The "agent" in this simulation was constrained by real-world locomotive engineering requirements, such as a maximum track grade of 1% and specific railway widths. This approach treats the landscape as a navigable environment, effectively simulating how an engineer might route a train around fine geometric obstacles.

The Unity-based NavMesh interface, where red and green zones indicate traversability based on engineering constraints like slope.

2. Deep Inverse Reinforcement Learning (IRL)

For long-range predictions where high-resolution 3D reconstruction is computationally prohibitive, a machine learning approach was employed:

  • Cost Map Generation: A U-Net convolutional encoder-decoder was used to transform 2D Digital Elevation Models (DEM) into "traversability cost maps."
  • Learning from Demonstration: The IRL model was trained on existing railway paths to learn the implicit "cost function" used by human engineers (e.g., avoiding steep inclines or deep valleys).
  • Policy Estimation: The model uses the learned cost map to predict the most probable future trajectory of the infrastructure project.

Datasets and Case Study

The primary test case was the Addis Ababa-Djibouti Railway in Ethiopia. The researchers utilized Maxar WorldView-3 imagery for high-resolution 3D surface modeling and NASA Shuttle Radar Topography Mission (SRTM) data (60m resolution) for the long-range IRL models. Access to the Maxar archive was facilitated through the EnhancedView web library.

Key Results and Insights

The study yielded several critical insights into the efficiency of directed monitoring:

  • Engineering Simulation: NavMesh tools proved mathematically robust enough to simulate real-world constraints, successfully routing a 10km test path while maintaining the 1% grade limit.
  • Data Efficiency: The IRL approach demonstrated that reasonable path predictions could be made using low-resolution 2D data. Analyzing the entire Djibouti-Addis Ababa range required less than 50MB of SRTM data, a massive reduction compared to the 500GB+ required for high-resolution photogrammetric analysis of the same area.
  • Selective Detection: By predicting the path, the required search area for high-resolution imagery is narrowed to a tiny fraction of the original request, saving significant bandwidth and cost.

Limitations and Future Work

While successful, the current model relies primarily on terrain relief. It does not yet account for material classification (e.g., soil type), which affects construction speed and cost. Future iterations aim to integrate soil and weather data to provide not just a predicted path, but a predicted construction timeline. Additionally, the team noted the difficulty in obtaining frequent "during-construction" satellite fly-overs for temporal validation.

Conclusion

By combining the geometric precision of game-engine pathfinding with the predictive power of Inverse Reinforcement Learning, this framework offers a scalable solution for infrastructure monitoring. It transforms the "needle in a haystack" problem into a directed search, enabling more efficient use of satellite resources in undeveloped regions.

Summary updated

Abstract

In the globalized economic world, it has become important to understand the purpose behind infrastructural and construction initiatives occurring within developing regions of the earth. This is critical when the financing for such projects must be coming from external sources, as is occurring throughout major portions of the African continent. When it comes to imagery analysis to research these regions, ground and aerial coverage is either non-existent or not commonly acquired. However, imagery from a large number of commercial, private, and government satellites have produced enormous datasets with global coverage, compiling geospatial resources that can be mined and processed using machine learning algorithms and neural networks. The downside is that a majority of these geospatial data resources are in a state of technical stasis, as it is difficult to quickly parse and determine a plan for request and processing when acquiring satellite image data. A goal of this research is to allow automated monitoring for largescale infrastructure projects, such as railways, to determine reliable metrics that define and predict the direction construction initiatives could take, allowing for a directed monitoring via narrowed and targeted satellite imagery requests. By utilizing photogrammetric techniques on available satellite data to create 3D Meshes and Digital Surface Models (DSM) we hope to effectively predict transport routes. In understanding the potential directions that largescale transport infrastructure will take through predictive modeling, it becomes much easier to track, understand, and monitor progress, especially in areas with limited imagery coverage.
BibTeX
@misc{mccullough2020utilizing,
  author = {Kyle McCullough and Andrew Feng and Meida Chen and Ryan McAlinden},
  title = {{Utilizing Satellite Imagery Datasets and Machine Learning Data Models to Evaluate Infrastructure Change in Undeveloped Regions}},
  year = {2020},
  eprint = {2009.00185},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2009.00185},
}
Figure