About this project

What this is

The biggest land changes for the area I grew up in: the Seattle metro and Mount Rainier. It uses a public satellite dataset that covers the whole planet at 10 meter resolution, every year back to 2017, a way to see what's actually been changing around us.

Right now it covers two areas as a starting point. If there's interest, more Washington areas are an easy add. See the How it works section below for why.

Spotted something that looks wrong, or want to see your own area covered? Email [email protected].

How it works

Every 10 meter patch of land has a satellite "embedding" for each year: a vector that summarizes what that patch looked like from space, derived from Sentinel-1, Sentinel-2 and Landsat imagery by Ai2's OlmoEarth model. For each area, two years of embeddings are compared pixel by pixel: 0 means "looks identical," higher means "looks very different."

High-change pixels are grouped into cells, then merged into "sites" (connected blobs of change) with a minimum size, so single noisy pixels don't make the list. Sites are ranked by area × how much they changed. A third, middle year is also compared, to help separate changes that stuck around from one-year flicker (a dry year, a cloudy pass, a crop rotation).

• Mount Rainier · 2017 to 2025 · 19 sites
• Seattle Metro · 2017 to 2025 · 81 sites

Limitations

• Early years (2017 to 2020) have incomplete coverage in some areas; each area here uses a year pair with full coverage.
• Change detection uses a local 20km-neighborhood z-score plus a 3-year persistence check (a site must still look different in a third, middle year), not a full crop-rotation / water-level-aware noise model.
• Categorization is a best guess and has not undergone rigorous evaluation.
• Place names come from reverse geocoding a site's centroid and may be approximate.
• Before/after imagery is a single Sentinel-2 scene, picked for the clearest view of each specific site rather than a cloud-free composite, so some haze is possible.