AI for image compression

The model was developed to demonstrate how onboard AI post-processing can reduce the amount of data that is sent to the ground without losing relevant information.  

Images are compressed on board. As input, the model accepts single-band images acquired by the multispectral camera. These images are compressed through the onboard AI encoder. The images are then reconstructed by the means of a dedicated decoder after they are downlinked to the ground. 

This application will first be demonstrated over Europe, focusing on building detection.  

During the development of the model, a real application scenario was used as a reference (semantic segmentation) in which the decompressed images were used to predict building masks along with the original acquisitions. The obtained masks were then compared to assess the differences between the two cases, confirming that the decompressed images can be used with the same results as the originals. 

The deep compression application is developed by GEO-K in Italy.