Deep learning from space – #ISD

Deep learning from space  -  #ISD

Sobolt is at the Industry Space Days on September 11 & 12

Viewing the whole world is only the first step in making sense of what we see. Extracting relevant information means analyzing all these data; a daunting prospect in many areas, even if we know when and where to look. Artificial Intelligence can lighten the load by taking many tasks off our hands. If you could tell a computer what you want it to see from above, what would that be?

Automating image interpretation

Sobolt is committed to automating large-scale remote sensing analyses. Using predominantly the optical part of the spectrum, we use state-of-the-art deep learning methods in monitoring, mapping, and data enhancement. Our integral deep learning approach allows for powerful yet flexible models, extracting relevant information from remote sensing data.

The vast amount of data in remote sensing makes it ideally suited for unsupervised deep learning

Our deep learning applications are not separate from one another, but benefit greatly from synergy across models. Unsupervised learning leverages the large amount of available remote sensing data to learn rich representations. These data representations are key to solving problems in computer vision, and can be used in various applications in different domains.

Super resolution

Enhance the spatial resolution to acquire realistic looking sharper images, up to 4x the initial resolution. Deep learning ensures that the upsampled data looks more natural than alternative methods. Data with a ground resolution of 10 cm can be significantly more valuable than the 30 cm counterpart.

Lower resolution satellite image before upsampling

Upsampled 'super resolution' image using a deep learning model

Object detection

Using higher resolution optical data, actual object detection becomes possible. Deep learning is optimally suited, because it allows one to learn abstract concepts, like cars, boats, planes and buildings. Many different types of objects can be learned to identify.

Deep change detection

Not all changes in remote sensing data are easy to spot, let alone establish their relevance. Moreover, what is relevant will change significantly per use case. Deep change detection not only learns an abstract representation of changes, but can classify their relevance and other attributes.

Get in touch

Whether you are looking to automate your monitoring or mapping capabilities, enhancing remote sensing and spatial data, or starting an innovative collaboration, we would love to discuss opportunities.

Click here to get in touch or look us up at ISD 2018!


Comments are closed.