Map satellite

Google Maps gets a new 700 trillion pixel cloud-free satellite map

Over a billion people use Google Maps every month, making it probably the most popular atlas ever created. On Monday, it gets a facelift, and its many users will see something different as they survey the planet’s forests, fields, seas, and cities.

Google has added nearly 1.5 trillion pixels of new data to its service*. The new map, which goes live this week for all Google Maps and Google Earth users, consists of newer, more detailed and higher contrast orbital images than the previous version.

More importantly, this new map has fewer clouds than before, this is only the second time that Google has unveiled a “cloudless” map. Google hadn’t updated its low- and medium-resolution satellite map for three years.

The improvements can be seen in the representation of Christmas Island on the new map. Nearly a thousand miles from Australia, the island was largely untouched by human settlement until the last two centuries. Its remoteness gives it a unique ecology, but given its location in the middle of the tropical Indian Ocean, it is often obscured by clouds. The new map eliminates them:

Google/Landsat

A 99-acre immigration detention center operated on behalf of the Australian government is now clearly visible; it is the only brown patch of development in the northwest “arm” of the island. In the old version of the map, the detention center was harder to distinguish from the clouds. Settlements on the east of the island are also now fully visible. Compare the old version of the map:

Google/Landsat

The new map also doesn’t include the darker diagonal lines that seem to slice through the old scene above. These lines were caused by a physical malfunction on Landsat 7, the US government satellite that provided imagery data for the old map. The new version of the map includes data from Landsat 8, the most recent version of the same satellite, allowing Google to eliminate ugly artifacts.

Landsat 8 has been operational since spring 2013. Its inclusion helps the new map to contain more recent data and, therefore, new structures. In Toyko’s new version of the map, Haneda Airport’s “D” runway can be seen in the bottom center of the frame. The stand-alone track was built on land reclaimed in 2010.

Google/Landsat

In the older version of the same scene, track D looks more translucent. The map as a whole is also blurrier, less saturated, and more generically gray.

Google/Landsat

This “ghostly” trail effect shows how Google is making its maps cloudless in the first place. None of the images above were captured in a single shot by a lone satellite, the way a camera might capture a snapshot. Instead, Google engineers used a recently developed mapping technique called mosaicking.

Mosaicking draws on the vast image archive created by the US government’s Landsat program, a series of satellites that have photographed the Earth’s surface every 16 days since the 1970s. (Landsat 7 and Landsat 8 are only the most recent of these devices.)

Taken individually, most of these images captured by Landsat sensors include some clouds. It’s logic: About 70% of the Earth’s land surface is covered with clouds every day. Over time, however, very few places are completely obscured by clouds. Mosaicking connects the cloudless parts through the power of surprising and elegant math.

A mosaicking algorithm inspects each imagery pixel individually, across all images of that particular pixel collected by Landsat 7 and 8. (If the archive is properly calibrated, that pixel should describe the same place on Earth no matter when it was taken.) Essentially, the algorithm takes an initial average color value for that pixel over time. Then it removes images that are much lighter than this average, because they likely include clouds, and averages the most recent set of remaining photos, now cloud-free, to find a final color value.

Then it runs this program for the next pixel. Eventually, these “best pixels” are assembled into a single map, a mosaic. The team analyzed more than 700 trillion pixels of satellite data throughout the mosaicking process.

Google’s version of this algorithm takes into account other special circumstances, such as seasonal effects. Matt Hancher, an engineer, called the process “the glorified calculation of the median”. The company’s Maps and Earth team has been using a mosaicking algorithm to generate its satellite map since June 2013.

A strange feature of the mosaicking process is that it captures the essential character of what a place looks like even though it actually encompasses years of data. Writing on a mosaic map of the UK produced by startup Mapbox in May 2013, journalist Tim Maly wrote: “At no time in UK history has it looked like this. [map]. Yet this [map] that’s exactly what it looks like.

Mosaicking is computationally intensive: it has only flourished in the past three years due to industry-wide improvements in cloud infrastructure. Google’s new satellite map has finally gobbled up more than a petabyte of imagery data. It took more than six million hours of computation, a processing feat that about 43,000 computers in Google’s cloud infrastructure handled in less than a week.

Google’s new update does not include images at the highest zoom levels, such as those needed to closely inspect a single-family home, swimming pool or baseball field. These images are not from Landsat at all, but from a mix of other public and private aerial and space cameras, including DigitalGlobe’s high-resolution satellites. This update only covers zoom levels from zero to 12.

Yet the new map still reveals a surprising amount of land change. Compare these images of the Columbia Glacier, an ice floe on the south coast of Alaska near Valdez. The first is from the 2013 version of Google’s cloudless map, and the glacier is visible despite Landsat 7’s diagonal stripes:

Google/Landsat

In the new version of the map, the diagonal scars are gone, but the glacier has also clearly retreated further inland:

Google/Landsat

The recentness of the new map reveals more than climate change. Compare these two images of Minneapolis. The first is from the 2013 version of the cloudless map.

Google/Landsat

The new version of the map is less green – the highway tans stand out more and the lakes are a darker navy. But the new map also reveals a long scar trail in the city’s northern core, left by the city’s deadly tornado in 2011.

Google/Landsat

Here’s that section enlarged, with the scar highlighted:

Google/Landsat

Image processing for this latest map was done entirely in Google Earth Engine, the company’s geospatial cloud infrastructure. In fact, the entire algorithm to create the cloudless map was written in Javascript in the Earth Engine development interface. The same interface was also used to calculate forest research and forecast malaria updates.

And, in some ways, Google Earth Engine is what’s ultimately the focus here. Terrestrial cloud computing is expected to be increasingly important for financial trading and business investment. The improved Earth Engine is part of why Google bought Terra Bella, a satellite company, two years ago. For big tech companies, how something got on the map will soon be as important as what’s there in the first place.


* This post originally stated that Google had added 700 trillion pixels of data to its satellite map layer. In fact, it analyzed 700 trillion pixels of data to create a final product of 1.5 trillion pixels. We regret the error.