Ryelore’s AI Platform
Artificial intelligence algorithms give best in class performance for analysing satellite images. However, applications are constrained by training data availability. Algorithm and object of interest complexity increases training data requirements while training data is limited by prevalence of objects of interest, re-visit rates, image archive and known objects of interest locations and high cost of imagery.
As a solution, we developed an advanced artificial intelligence platform that analyses satellite imagery. We address training data limitations with (1) iterative geo-scans where an initial neural network algorithm is built and repeatedly applied to the target geography in order to find new training samples that further improve algorithm performance and (2) generative adversarial networks (GANs) that generate fake images of an object of interest to augment training data.
During the development of the artificial intelligence platform, we demonstrated how a state of the art semantic segmentation algorithm is developed and fine-tuned on a challenging use case where i) region of interest was 9M square km and large, ii) the object of interest showed a great of amount of diversity, iii) landscapes, illumination and cloud cover further increased variation and iv) training data was limited as location of many object of interest was unknown. We showed how a state of the art GAN algorithm is trained to generate fake objects of interest
that are indistinguishable from real ones. We also demonstrated that such an artificial intelligence platform could
deliver excellent results to address a use case that could provide unique data results.
We have developed a re-usable artificial intelligence platform that analyses satellite imagery. It enables development of new and challenging use cases at a lower cost through the application of advanced machine learning algorithms and software components that manage large satellite imagery datasets.
We will continue developing the state of the art artificial intelligence platform, adding new types of neural networks, making management of large satellite image datasets easier and addressing new commercial and scientific use cases.