Although the operational production of Theia’s OSO map was transferred to CNES’ MUSCATE team, at CESBIO we keep on working on the improvement of the algorithms ans the tools (the iota2 processing chain).

We have recently implemented into iota2 a new contextual classification method developed by Dawa Derksen during his PhD (the slides of his oral defense are available here). The details of the method have been published in this paper, where he shows that it performs as well as Deep Learning approaches, but with much lower computational (and energy) costs.

As most of the researches who publish in technical journals, we could have stopped there, but Dawa wrote code using the Orfeo Toolbox and made it free software (public money, public code), then Arthur Vincent (iota2‘s guru’) made everything needed so that iota2 can call this code and use it to process entire countries!

See below this small animation comparing the results of the « classic OSO » map and the contextual version.


Aerial, classic, contextual

In order for you to make your own opinion, we have made available for download the map over the southern 1/3 of Metropolitan France (around 30 Sentinel-2 tiles). The GeoTiff file is available here:

For reminders, the classic OSO version is available here:

You will be able to compare both versions and give us your point of view. The 2019 map is on its way, but it will not use the contextual version of the processing chain. We need users’ feedback to decide whether this version is preferred to the previous one. If this was the case, upcoming products could use the new algorithm.

Don’t hesitate to send us your feedback.


One thought on “Contextual classification for Theia’s land cover (OSO) map

  1. It looks like the contextual classification yielded less noises! Very cool! Is it possible to have the map covering the whole France?


Laisser un commentaire

Votre adresse e-mail ne sera pas publiée.

Ce site utilise Akismet pour réduire les indésirables. En savoir plus sur comment les données de vos commentaires sont utilisées.