AI-Driven Urban Monitoring for Yerevan

Funded by: Afeyan Foundation

PI: Hayk Nersisyan

Researcher: Hovhannes Grigoryan, Nerses Srapionyan

This project develops a low-cost, reproducible AI pipeline for monitoring urban development in Yerevan and its surroundings using satellite imagery and geospatial machine learning. The city is divided into approximately 65-meter H3 hexagonal cells, where each cell represents one satellite image patch.           For each patch, we extract self-supervised visual embeddings using DINO-based models, allowing the system to learn meaningful land-use patterns without requiring large manually labeled datasets.

Important results so far:

We have completed the first end-to-end version of the unsupervised discovery pipeline for Yerevan. The study area was partitioned into 28,832 H3 hexagonal cells, each corresponding to one satellite image patch. Using self-supervised DINO embeddings followed by UMAP dimensionality reduction and HDBSCAN clustering, we identified 10 visually coherent clusters with zero manual labels.

Initial visual inspection shows promising differentiation between several meaningful urban and peri-urban patterns, including dense residential areas, open or sparse land, and industrial zones. These results suggest that the model is already capturing useful spatial structure from satellite imagery without manual annotation. Spatial validation is still ongoing, and the next phase focuses on 2019–2025 change detection and Markov-chain-based transition modeling to study how Yerevan’s urban form is evolving over time.

Figure 1. AI-driven urban monitoring workflow for Yerevan. Open satellite imagery is divided into approximately 65-meter H3 hexagonal cells; each cell is processed as an image patch using self-supervised DINO embeddings; UMAP and HDBSCAN are then used to discover visually coherent clusters. Early results over Greater Yerevan identify 10 clusters across 28,832 cells without manual labels, with examples corresponding to dense residential, open/sparse, and industrial areas.

The next technical steps are spatial validation of the discovered clusters, integration of 2019 imagery for change detection, and development of Markov-chain transition analysis to quantify how land-use patterns evolve over time.