Exploring the potential of AlphaEarth and TESSERA embeddings for Fine-scale Local Climate Zone Mapping: A case study across five cities in Switzerland

Imagine you are trying to draw a map of a city, but instead of just showing where the parks and buildings are, you want to show exactly how the city "feels" to the weather. Does a street feel like a hot oven because of tall, tight buildings? Or does it feel like a cool breeze because of scattered trees? In the scientific world, this is called mapping Local Climate Zones (LCZ).

For a long time, scientists have had a "rough draft" of these maps. Think of it like a low-resolution photo where every pixel is the size of a football field (100 meters). It's useful for a big picture, but if you want to understand the climate of a specific neighborhood or a single street, that photo is too blurry.

This paper is like a team of scientists trying to turn that blurry, football-field-sized photo into a crisp, high-definition image (10 meters per pixel) for five major Swiss cities: Bern, Basel, Geneva, Lausanne, and Zurich.

Here is how they did it, using some creative analogies:

1. The Three "Eyes" Looking at the City

To create these high-definition maps, the researchers didn't just look at the city with one pair of eyes. They tested three different "super-vision" tools to see which one could best translate a blurry reference map into a sharp one:

• The Classic Eye (S1S2): This is the traditional method. It looks at the city using standard satellite photos (Sentinel-1 and Sentinel-2). It's like looking at a city through a clear window, but you have to do a lot of cleaning (removing clouds, adjusting colors) before you can see anything clearly.
• The "Patchwork" Eye (AlphaEarth): This is a new AI tool from Google. Imagine it doesn't look at the city pixel-by-pixel, but rather looks at small "patches" or neighborhoods at a time, learning the texture of the whole block. It's like a quilt maker who understands the pattern of a whole section of fabric rather than just one thread.
• The "Time-Traveler" Eye (TESSERA): This is another new AI tool. Instead of just looking at a snapshot, it looks at the city's "movie" over the whole year. It understands how the city changes with the seasons—when trees turn green, when snow melts, and how the ground looks in summer versus winter. It creates a "time capsule" of data.

2. The "Smart Painter" (The AI Model)

The researchers used a special type of AI called an Attention U-Net. Think of this AI as a very smart painter who is given a blurry, low-resolution sketch (the reference map) and asked to paint a high-definition masterpiece.

The "Attention" part is like the painter wearing special glasses that help them focus on the most important details. If the painter sees a cluster of tall buildings, the glasses tell them, "Pay extra attention here; this is a dense city zone." If they see a river, the glasses say, "Focus here; this is water."

3. The Three Experiments

The team ran three different tests to see how well their "Smart Painter" worked:

• Experiment 1: The Multi-City Challenge. They trained the painter to recognize all five cities at once using the "football field" (100m) reference maps.
  • Result: The Time-Traveler Eye (TESSERA) was the best painter. It produced the sharpest maps. The Classic Eye was okay, but the Patchwork Eye (AlphaEarth) struggled a bit more with the variety of different cities.
• Experiment 2: The High-Resolution Test. They focused only on Bern, where they had a slightly sharper reference map (78m instead of 100m).
  • Result: With a better sketch to start with, all painters got better. The Time-Traveler Eye (TESSERA) still won, but the Patchwork Eye (AlphaEarth) caught up and did almost as well. The Classic Eye was still the weakest.
• Experiment 3: The Time Travel Test. They trained the painter on 2024 data and asked it to paint the 2025 city without retraining.
  • Result: This was tricky. The city changes slightly every year (trees grow, weather shifts). The Classic Eye (S1S2) was the most stable; it didn't panic when the year changed. The Time-Traveler Eye (TESSERA) did well but was a bit more sensitive to the changes. The Patchwork Eye (AlphaEarth) got confused the most when the year changed.

4. The Big Takeaways

• The Magic of "Embeddings": The study found that using these new AI tools (TESSERA and AlphaEarth) is like having a pre-packaged "smart kit." You don't need to spend weeks cleaning and preparing the raw satellite photos yourself. The AI has already done the hard work of understanding the data.
• TESSERA is the Star: The tool that looks at the city's "seasonal movie" (TESSERA) consistently created the most accurate maps, especially when looking at different types of cities.
• The Bottleneck: Even with the best tools, the quality of the final map depends heavily on the quality of the "sketch" (the reference data) they started with. If the reference map is blurry, the final painting will have limits. The researchers found that having a sharper reference map (like the 78m one for Bern) made a huge difference.
• Time is Tricky: While the new AI tools are powerful, they are sensitive to the time of year. A model trained on a summer city might get confused if shown a winter city, unless it is specifically taught to handle those changes.

In Summary

This paper proves that we can use advanced AI "super-eyes" to turn blurry, old climate maps into sharp, detailed, street-level guides. The best tool so far is TESSERA, which understands the city's seasonal changes. While there are still challenges with how the maps handle different years and rare city types, this approach opens the door to creating detailed climate maps for cities all over the world without needing to send people out to measure every single street.