Abstract
Abstract. Cubesats platforms expansion increases the need to simplify payloads and to optimize downlink data capabilities. A promising solution is to enhance on-board software, in order to take early decisions, automatically. However, the most efficient methods for data analysis are generally large deep neural networks (DNN) oversized to be loaded and processed on limited hardware capacities of cubesats. To use them, we must reduce the size of DNN while accommodating efficiency in terms of both accuracy and inference cost. In this paper, we propose a distillation method which reduces image segmentation deep neural network’s size to fit into on board processors. This method is presented through a ship detection example comparing accuracy and inference costs for several networks.
Highlights
Nowadays, cubesats platforms appear to be an attractive and lowcost tool to acquire image data from outer space
The F2-score is the metric used in the original ADS challenge and, since ground truth of the test set is not available, evaluation on the test set is done through Kaggle website which provides the F2-score as results. It is computed based on the number of true positives (TP), false negatives (FN), and false positives (FP) resulting from comparing the predicted object to all ground truth objects
Distillation is a generic and re-usable workflow for simplifying deep learning (DL) networks for defining new Earth Observation (EO) products generated on-board the satellites
Summary
Cubesats platforms appear to be an attractive and lowcost tool to acquire image data from outer space. Downlink data capabilities are the bottleneck of the cubesats platforms. The compression ratio required for image payloads is too high to be obtained by existing compression methods (Buciluǎ et al, 2006; Frosst and Hinton, 2017; Hinton et al, 2015). On-board, feature extraction based on deep learning (DL) (Greenland et al, 2018) provides, the most efficient solution for upstream data reduction. It can provide high-value information directly exploitable. The most efficient DL models are usually cumbersome due to ensembling methodologies, e.g. hundreds of millions of parameters in the case of the winners of Airbus Ship Detection challenge proposed on Kaggle, and are not compatible with the limited performances of the processors available on board
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
