GeoVisual Search from Descartes Labs makes the Earth searchable

Scavenger hunts just got significantly more tactical with Descartes Labs’ new GeoVisual Search. Finding shipping containers, runways and even parking lots on a global scale is no problem with the free tool made available today. Descartes Labs is a geospatial analytics startup based in Los Alamos, New Mexico. The company specializes in analyzing satellite imagery and other global data with machine learning to power predictive analytics for agriculture and other key industries.

Users can scour the earth’s surface by placing a provided bounding box around any object they would like to search for. GeoVisual Search returns other instances of the same object across the world. The team is still tinkering so it’s not able to return an exhaustive list of every occurrence of a given feature.

But such a feature would require a more advanced search. The models powering GeoVisual Search don’t have the innate knowledge to accurately differentiate between windmills and other turbines. Descartes Labs CEO Mark Johnson hints that this could be accommodated with a Tinder-like interface, so users could swipe right and left on returned images to fine-tine their search.

Descartes isn’t charging for GeoVisual Search, though Johnson alluded to a future paid version for analysts. That more complete platform could include the additional aforementioned features.

Large tech companies have the resources today to explore data in ways that the average researcher cannot. Facebook did a population density analysis by using computer vision to analyze structures. The company hopes to use the data to inform its global connectivity efforts. GeoVisual Search could eventually make it possible for groups with fewer resources to undertake similar projects.

“We’re looking for customers to come in who want to do complex jobs,” said Johnson.

In the background, the company is running a 50-layer ResNet built with Keras and pre-trained on ImageNet. From there, it was about making optimizations for searching satellite imagery, like fine-tuning for classification of designated OpenStreetMap objects.

“There is no religion around the algorithm,” added Johnson. “We didn’t need deep learning for everything.”

Descartes Labs leaned heavily on composite imagery to build its image search. To improve the quality of its image corpus, the company has been layering satellite maps on top of each other and optimizing for the best pixels across large areas. This is in addition to the traditional cleaning and atmospheric corrections that have to be made to cut down on image noise.

The overall project builds on Terrapattern, an effort by Carnegie Mellon to build a similar search tool. But the key difference is that Descartes’ GeoVisual Search combines NAIP Arial Imagery, PlanetScope and Landsat 8 to allow the search engine to query the entire globe. Terrapattern is solely focused on a handful of cities, though its search is quite accurate.