Skip Navigation
Center Independent Research & Development: GSFC IRAD

An Innovative and Modular Approach for Deep Learning for Constellations of Smallsats

Completed Technology Project

Project Introduction

The primary research and development goal of this research is to enable the usage of advanced machine learning algorithms, particularly deep learning, on embedded hardware systems such as satellites. This work would allow for powerful classification engines to be run on modest hardware, and lead the way for constellations of small satellites to perform near real-time triage of interesting events on the ground.

The primary long term objective of this project is have a machine learning framework that is geared towards being used on highly constrained hardware, and to show that framework being used in a simulated distributed spacecraft application. In the short term, the research will focus on one particular application of deep learning as a proof of concept. In order to do this, the first objective will be to curate a dataset that will be used as a training set for all permutations of the deep learning artificial neural network (ANN). The next objective will be to determine an optimal architecture for the ANN, and then to implement it in software for testing. Finally, once an optimal architecture is chosen, it will be mapped to hardware using energy efficient FPGAs.

The secondary objective is to use the previously developed ANN to do scene classification on real data, such as the AVIRIS, MODIS, and Hyperion multispectral data sets. The ANN will be trained to look for interesting scenes in a vast amount of data, and can be configured a number of different ways in order to look for fires, floods, algal blooms, and other situations. 

The final milestone will be a full definition of a prototypical mission that can be used as an example of the proper use of the framework. This will include a generalization of the framework, which means that more cores will be added to accommodate a wider range of ANN architectures. Additionally, a mission will be described for using the framework to do near real-time detection of wildfires by analyzing multispectral data with the ANN developed in the previous milestone. This milestone is the final one for FY17 and will conclude in September.

More »

Anticipated Benefits

Primary U.S. Work Locations and Key Partners

Share this Project

Organizational Responsibility

Project Management

Project Duration

Technology Maturity (TRL)

Technology Areas

Target Destinations

Light bulb

Suggest an Edit

Recommend changes and additions to this project record.