Search the Planet with AI
The satellite image above shows an area of recent deforestation in the Amazon. I found this example using a new satellite imagery search tool from Planet, which allows you to search the Earth using nothing more complicated than a natural-language description.
Machine learning has been used to search satellite imagery for patterns and geographical features for quite some time. Around a decade ago projects such as Terrapattern and Descartes Labs' GeoVisual Search (both now in the dead pool) demonstrated how computers could scan enormous collections of satellite images to find visually similar features.
Now Planet Labs has brought this idea back in a rather more ambitious form with Planet AI. Its new application allows you to describe what you are looking for in ordinary language and then searches a global index of satellite imagery for matching locations.
Instead of having to know the coordinates of somewhere you want to investigate, or manually scan satellite images looking for something interesting, you can simply ask Planet AI to find it for you.
Searching the Earth
Planet AI uses machine learning to create a searchable global index of visual features extracted from Planet's monthly satellite mosaics. You can enter queries such as "recent deforestation in the Amazon", "large solar farms in the UK" or "open-pit mining activity" and the application will attempt to locate matching examples.
Once it finds potential matches, you can explore their locations on a map and investigate more than a decade of historical satellite imagery. This makes it possible to compare locations over time and create timelapses showing how landscapes have changed.
There is also an Analyze mode, which uses AI agents to undertake more complex investigations. It can combine satellite imagery with publicly available web information, incorporate uploaded images and CSV files, and even set up scheduled monitoring reports.
Planet AI is currently free to access after registration, although users receive a limited allocation of free credits. Planet doesn't appear to have published an official refill schedule (daily or monthly) for these free starter credits - so this may be a limited time offer of free credits.
A question of resolution
There is an important limitation to Planet AI's visual search: the resolution of the imagery being searched.
PlanetScope satellite imagery has a resolution of approximately 3 metres per pixel. This is excellent for observing large geographical features and changes in the landscape, but considerably less useful when trying to identify small individual objects.
I tried searching for swimming pools in London. Unfortunately, the results were almost entirely false positives. The application returned locations that were clearly not pools and failed to identify any actual outdoor swimming pools or lidos.
This isn't particularly surprising. A typical 8 × 4 metre swimming pool occupies an area of only a few pixels in a 3-metre-resolution image. There simply isn't enough visual detail to reliably distinguish a swimming pool from a blue roof, a covered structure or some other similarly coloured feature.
Planet AI is therefore likely to work best when searching for large geographical features with distinctive visual characteristics.
Try searching for:
Large solar farms and wind farms.
Open-pit mines and quarries.
Extensive areas of deforestation.
Large industrial estates and warehouses.
Ports, airports and container terminals.
Reservoirs and wastewater treatment facilities.
You are much less likely to have success searching for individual cars, swimming pools, small buildings or other objects occupying only a handful of pixels.
Exploring the Earth
Planet AI is an interesting development in the use of machine learning to understand satellite imagery. Its greatest appeal is that it removes much of the technical knowledge traditionally required to explore enormous collections of Earth observation data.
You no longer need to know where to look before you start looking. You can begin with an idea, describe it in ordinary language and let the AI suggest places worth investigating.
Via: Webcurios

Comments