Discovering Earth's Twin Landscapes: How Similar Earth Uses AI to Find Matches

Similar Earth is an open-source tool that lets you drop a pin anywhere on the planet and instantly find every other location that shares the same visual and environmental signature. Powered by Google DeepMind's AlphaEarth foundation model, it creates a global heatmap of look-alike places in about two seconds. Below, we explore how this works behind the scenes, from the challenge of planetary-scale computation to the clever caching strategies that make it possible.

What exactly does Similar Earth do?

Similar Earth is an interactive heatmap engine. You drop one or more reference pins on any location — a vineyard in Mendoza, a solar farm in the Mojave, or a mangrove forest in Bangladesh — and within seconds it returns a world map color-coded by similarity. Every 10-meter patch on Earth is compared to your pin, and areas that share a similar environmental signature glow brighter. For example, a pin on a coffee farm in Colombia lights up the highlands of Ethiopia, Vietnam, and Costa Rica — all places where coffee grows for the same underlying climatic and soil reasons. The heatmap itself is the core product, and the entire system is optimized to deliver it fast and cheaply at any zoom level.

How does the system find look-alike places so quickly?

The secret lies in three clever moves: a coarse grid kept in memory, on-demand refinement at high resolution, and aggressive caching. First, the planet's land surface is divided into 2-kilometer squares (instead of the full 10-meter resolution). This coarse grid contains only about 40 million pixels, and after compressing each location's 64-dimensional vector into 8-bit integers, the entire dataset fits in just 2.6 GB of RAM — comfortably on a standard cloud server. When you drop a pin, the server computes a max-pooled dot product between your reference vector and each pixel in parallel. The result is a low-resolution heatmap in under a second. For zoomed-in views, it refines only the visible area at full 10-meter resolution using precomputed embeddings stored on disk.

What role does AlphaEarth play?

Similar Earth would be impossible without AlphaEarth Foundations, a geospatial AI model released by Google DeepMind in 2025. AlphaEarth processes years of satellite imagery, climate data, terrain, and seasonal variations to produce a compact 64-number "embedding" for every 10-meter patch on Earth. Two patches with similar embeddings share a similar environmental signature — meaning similar vegetation, climate patterns, soil types, and land use. The beauty is that finding matches worldwide reduces to a simple mathematical operation: a dot product between those 64 numbers. No complex simulation or image comparison is needed; just a fast vector similarity search.

Why can't you just precompute every possible heatmap?

Computing a 10-meter resolution heatmap for the entire planet on every request would be impossibly slow and expensive. Earth's land area is roughly 150 million km², which at 10-meter resolution means about 1.5 trillion pixels. Doing a dot product for each of those pixels every time would take hours and cost a significant fraction of a dollar per query. Precomputing all possible heatmaps ahead of time is also infeasible because the number of possible reference pin combinations is effectively infinite. The solution, as outlined in the three moves, is to trade off absolute accuracy for speed at the coarse level, then refine only where needed.

How does the in-memory grid work in practice?

At startup, the server loads the entire 2-kilometer resolution grid into a single array: var grid [40_000_000][64]int8. This 2.6 GB structure stays resident, always ready. When a user drops a pin, the server processes all 40 million pixels in parallel using worker shards. The code computes a max-pooled dot product — for each pixel, it finds the best similarity score among all reference pins (if multiple are dropped) by summing element-wise products over the 64 dimensions. The result is an array of scores that's rendered as a global heatmap. Because the grid is quantized and cached aggressively, the whole operation takes roughly two seconds. For higher zoom levels, the server fetches fine-grained embeddings from disk only for the visible map area, keeping response times low.

What are the main benefits of this approach?

The hybrid strategy delivers three key advantages:

• Speed: Most queries return a global heatmap in under two seconds, thanks to the always-hot coarse grid and parallel computation.
• Cost efficiency: By keeping only 2.6 GB of data in RAM and using cheap int8 quantization, the system runs on modest cloud instances without breaking the bank.
• Scalability: Aggressive caching and on-demand refinement mean the system can handle arbitrary zoom levels and any number of simultaneous users without proportional resource increases.

In short, Similar Earth proves that planet-scale similarity search is not only possible but practical with clever engineering and modern AI embeddings.