OpenScience 2017

Magellium was present at ESRIN, Frascatti, Italy on monday 09/10 to share its point of view over some use case in segmentation for the sentinel 2 imagery. This conference was a great opportunity to talk about upcoming challenges in data science, processing and distribution for many sentinel cases. Our presentation can be watched here around 1:01:00.

Abstract :

The arrival of sentinel 2 data provides a great opportunity for researchers and scientist to test their algorithms on a large scale with many variations, at least both in time and locations. Many of these methods originate from the computer vision and pattern vision communities and address detection, classification and recognition problems. The available resolutions (10 meter GDS for panchromatic image) forbid the search and monitoring of small objects such as cars for example. However, many challenges remain for agricultural and urban related problems. In this study, we have focused on the cloud detection, although it is partially addressed by the ground segment for level 1C product, we have found that was still an important matter of interest for the strengthening of the product quality, leading to more robust higher level object extractions. In this regard we also have focused on another recognition problem which deals with the building footprints recognition and extraction. Thus, this paper sums-up our experience with both of these problems using a few convolutional neural networks from the literature. Many companies have open sourced their framework, allowing at little cost the use of very powerful tools to enable the design, training and application of such networks both locally and in the cloud. From our experience, Tensor Flow with Keras is a good way to prototype networks. None the less, we emphasize that a huge work is required in building proper the data sets. As far as we are concerned, this task remains one of the most challenging phase in the whole process, since a high ground truth quality is often required to reach high quality detection rates or high quality segmentation masks. In this regard, we address some comments with respects to the available open ground truth data and elaborate on our results while broadening the discussion to the benefits of active learning and unsupervised learning.