Keywords: AI, Artificial Intelligence, Cloud Detection, Computer Vision, Hyperspectral, Neural Network, NN, On-Board, Sentinel

Nowadays, the greatest limitation to the adoption of the very high-resolution imager is the limited downlink capabilities. This problem is more evident in the case of micro and nano satellite systems.

AI has revolutionised Earth Observation methodologies, simplifying and changing the way data are processed and treated. To this aim, IngeniArs presents the SkyArt CNN: an extremely lightweight, fast, and accurate segmentation model. The SkyArt CNN allows classifying with 91% of accuracy each pixel of the input image into two classes: cloudy or not cloudy.

Thanks to its input-independent architecture, it is possible to feed the network with 8 bands of multispectral images with different sizes.

The network can be executed on different embedded platforms, from dedicated FPGA hardware accelerators to most common COTS hardware devices.

With a sufficient amount of data, it is possible to re-train the network for:

  • New on-board hardware;
  • Fine customization (more input layers, bigger input size, etc.);
  • Dedicated application for data analysis.

Key features

  • Segmentation network
  • 64KB memory footprint
  • 100 ms inference time
  • Available for GPU and Myriad-2 VPU
  • 3 input channels: RGB; Hyperspectral channels
  • Inference time related to the input image e.g. 192x192: 102ms
  • 91% of accuracy (1% of FP)

Other information



  • Neural Network Design
  • Artificial Intelligence for Space Embedded Systems


SkyArt Brochure

Call us at 0039 050 6220532 or email us at