Methods and tools for making deep learning-powered techniques robust

Making deep networks more robust against noise

With the current sensor advancements, hyperspectral satellite imaging is becoming a mature technology which captures a very detailed spectrum of light for each pixel.

Hyperspectral image analysis is commonly performed in a setting where neither the environment nor the hardware poses an obstacle.
However, if the automated analysis is to be performed on embedded hardware, e.g., on-board a satellite, constraints in electrical power or computational resources can lead to the degradation of their performance, as well as to an increase in their inference time.

Importantly, real-life satellite data is often affected by noise due to different real-life limitations which makes it even harder to process.

In BEETLES, we tackle such issues and strive to implement, evaluate, and integrate robust deep neural networks for effective hyperspectral image analysis, including classification, segmentation, unmixing, feature selection and much more, that are ready to be deployed in the wild.