Agentic Engineering for Space Data

Data Center Site Growth in the US

This tracker was built by an agent on Tilebox: workflow, code, and data, end to end.

Monitor data center construction across the United States using before-and-after Sentinel-2 satellite imagery. Each site is ranked by visible construction change, including new ground disturbance, structural growth, and other buildout signals detected between 2024 and 2026.

What is Tilebox?

Tilebox is an agentic development framework that enables AI coding agents to build and operate Earth Observation workflows. Agents work against the same operational context an engineer would: knowing what data is available, which steps ran, where something failed, and what came out. That visibility makes the agent's work something you can inspect and trust, instead of a black box.

Watch us build it View Agent Thread View GitHub
Filter
View

2 year data center construction signal

Comparison period: June 2024 - June 2026

How is the score computed

8 shown / 22 hidden

Change score

0 100

A high score means visible buildout changed.

The workflow compares a low-cloud Sentinel-2 scene before the target period with one after it. It ranks sites where valid pixels changed, the change looks construction-like, and the evidence is coherent enough to rise above noise.

01

Construction

Looks for built-up gain, exposed soil, vegetation loss, and brightness change. The workflow favors construction-like pixels over generic image difference.

max(area, fraction, direction)
02

Object

Checks whether changed pixels form connected components. A large pad or staging area should outrank scattered one-pixel noise.

connected changed area
03

Embedding

Uses Clay foundation-model patch embeddings to compare before and after crops. Patch distance catches semantic visual change beyond simple spectral thresholds.

1 - cosine similarity
04

Structural

Compares image structure across RGB and false-color views using SSIM (Structural Similarity Index). It compares luminance, contrast, and local structure between the before and after images.

1 - mean(SSIM)
Change Score: Weighted component scores

A 0-100 ranking that favors sites with strong construction evidence and supporting spatial or model agreement.

The four displayed subscores expose the evidence channels behind the ranking. They are not averaged equally: construction evidence drives the final score, while object coherence and Clay embedding agreement help confirm that the change is spatially meaningful. Structural change is shown as an additional image-similarity diagnostic.

Metrics

changed_area
Total changed footprint, reported as hectares and as percent of valid image pixels.
changed_construction_area
Construction-like changed footprint, also reported as hectares and as percent of valid image pixels.

Example: 59.12 ha changed, 34.80 ha construction-like, 6.59% changed pixels, 3.88% construction-like pixels.

Data center construction monitoring FAQ

Short answers for how this tracker finds construction change in satellite imagery.

How can satellite imagery monitor data center construction?

Satellite imagery can reveal construction signals such as cleared land, exposed soil, new building pads, roof growth, road changes, and staging areas. This tracker compares before-and-after Sentinel-2 scenes to identify visible changes around known US data center sites.

What does the change score mean?

The change score ranks sites by visible construction-like change between the comparison scenes. Higher scores indicate stronger evidence of buildout activity, based on construction pixels, connected changed areas, structural similarity, and model embedding distance.

What satellite imagery is used?

The tracker uses Sentinel-2 imagery, which provides repeat observations at roughly 10 meter resolution. The workflow selects low-cloud before-and-after scenes for each site and compares changes across the target period.

Can this detect every data center construction project?

No. The tracker detects visible surface change in available satellite imagery. Cloud cover, tree canopy, subtle interior work, small changes, or already-completed sites can reduce visible signal.