Publication of On-board Clustering Analysis

A paper authored by one of the masters students whom I co-advised (and co-authored by myself) has recently been published.

We investigated the possibility of using a shallow neural network, called a self-organizing map, as a way of clustering hyperspectral data on-board the HYPSO cubesat.

Clustering is a very flexible operation in machine learning which can pre-process data so that it can be analyzed by other algorithms more easily. For example, because clustering groups similar pixels together, a classification algorithm can operate more quickly by analysing the clusters (a few hundred) rather than the individual pixels (a few hundred thousand), which provides a significant speed improvement.

In itself, clustering can be used as a form of compression, as a summary of the data. We found that by training the SOM neural network on the ground before subsequently uploading the network weights to the satellite, this form of clustering leads to significant reductions in the amount of data while losing very little information. Most importantly, it seems feasibile from the HYPSO operational standpoint.