Monitoring war destruction from space using machine learning

Existing data on building destruction in conflict zones often rely on eyewitness reports or manual detection, making them scarce, incomplete, and potentially biased. This shortage of reliable data limits media reporting, humanitarian relief efforts, human-rights monitoring, reconstruction planning, and academic research.

This article presents an automated method of measuring destruction in high-resolution satellite imagery using deep-learning techniques enhanced with label augmentation and spatiotemporal smoothing to interpret the structure and progression of destruction.

As a proof of concept, the authors apply the method to the Syrian civil war, reconstructing destruction dynamics across major cities. Their approach produces destruction data with unprecedented resolution, frequency, and accuracy—making such information far more useful for researchers and practitioners.

Methodological Contributions

1. Label augmentation: A new strategy for expanding destruction class labels using contextual assumptions and additional information.
2. Two-stage classification process: Designed to smooth spatial and temporal noise inherent in high-resolution satellite data.

The model uses convolutional neural networks (CNNs) trained to detect features of destruction caused by heavy weaponry (e.g., artillery, bombing), such as rubble or bomb craters. The approach significantly improves the timeliness, resolution, and reliability of destruction data in conflict settings.