Satellite imagery promises to play an essential role in creating digital twins that could help model carbon emissions, increase farming yields and monitor the oceans. One hope is that the rising volume of satellite data could automatically detect changes, predict trends and recommend actions.
However, much of this data, in the form of multispectral imagery, lidar and radar data, is outside the realm of typical human experience, which makes it hard for humans to make sense of and label data for training better artificial intelligence (AI). Rendered AI hopes to fill this gap by combining its synthetic data workflow platform with new geographic information systems (GIS) industry partnerships.
A new partnership with GIS leader Esri will make it easier to combine satellite imagery with 3D content to generate example training sets. Another partnership with Rochester Institute of Technology’s Digital Imaging and Remote Sensing (DIRS) Laboratory will make it easier to combine the lab’s DIRSIG satellite synthetic data tools with Rendered.ai’s cloud-based platform for high-volume synthetic data generation.
“As we move away from images that are easy for humans to interpret, it becomes much harder to build labeled datasets that can be used in artificial intelligence training,” said Nathan Kundtz, cofounder and CEO of Rendered. “This has historically meant that these sensor types have largely been frozen out of the AI revolution.”
Synthetic data representing various kinds of satellite data will make it easier to train AI to automatically detect when farm yields drop, CO2 levels rise and coral reefs suffer.
Synthetic data superpowers
The company was founded in 2019 by Nathan Kundtz, Kyu Hwang, Duane Harkness and Ethan Sharratt. Kundtz said he conceived of the idea for the company after frustration with the lack of accurate sensor data for training AI while working in the satellite communications industry.
Real sensor data is biased toward common events that are easy to detect or collect, Kundtz said.
“Synthetic data offers the opportunity to overcome bias and gaps in real sensor data and is relatively inexpensive, encouraging innovation and experimentation,” Kundtz said.
The company has focused on developing a full infrastructure stack to integrate physics-based synthetic data into AI training workflows. This includes 3D model sourcing, simulation tools, content management, managed compute, application deployment, tasking, data provenance, annotations and data quality assessment. The platform allows computer vision engineers, synthetic data engineers and domain experts to collaborate, iterate quickly to generate data, run scenarios and drive toward improved AI performance.
“Our job has been to give those engineers superpowers — giving them the tools needed to become synthetic data engineers,” Kundtz said.
Overcoming GIS challenges
Kundtz says the GIS community is also desperate for automation because far more valuable data is being collected than humans can process. And synthetic data workflows for the GIS community can be more challenging than other domains.
“What makes GIS hard is that the range of things which might need to be simulated is incredibly large – literally the entire planet at many scales and the range of sensor types used is also highly diverse,” Kundtz explained.
For instance, much of GIS analysis uses sensors that operate outside what the human eye can see – like infrared, radio frequency and synthetic aperture radar sensing.
Customers are generating synthetic satellite imagery data for rare object detection, identifying land cover segmentation and building detectors for asset conditions. Others are using the tools to improve the ability to extract data from drone imagery, video and other geospatial content types.
As this volume of synthetic data continues to grow, automation offered by companies like Rendered AI offer the potential to apply AI to large datasets to solve real-world problems.