Semma-Drone is a collaboration project to map sea turtles in foraging areas thanks to drones and vision software. This presentation was made for the 36th international symposium on sea turtles (ISTS36).
1. Gwénaël Duclos – WIPSEA
ISTS36 Lima - Peru
Thursday 3rd of March 2016
Romain Dambreville – WIPSEA
Katia Ballorain – Parc Naturel Marin de Mayotte
David Corman – Agence des Aires Marines Protégées
Gwénaël Duclos – WIPSEA
Stéphane Ciccione, Claire Jean – Kélonia
Marc Lennon - ACTIMAR
2. Why ?
• Also on foraging areas
• Sea Turtles need efficient protection
• Not only on nesting beaches
> Requires abundance estimates of the population
3.
4. Hodgson, A., Kelly, N., & Peel, D. (2013). Unmanned aerial vehicules
(UAVs)for surveying marine fauna
Mejias, L., Duclos, G., Hodgson, A., & Maire, F., (2013, September).
Automated marine mammal detection from aerial imagery
Maire, F., Mejias, L., Hodgson, A., & Duclos, G. (2013, November).
Detection of dugongs from unmanned aerial vehicules.
5. SEMMA-DRONE project
Goals:
• Reduce costs
• Increase surveys frequency
Partners:
Suivi et Etude de la Mégafaune MArine par DRONE
Marine Magafauna Survey with DRONE
• Accelerate global mapping process
16. Applications
• At Mayotte MPA, these data are used to
evaluate:
– Population size
– Foraging habitat use
– Turtles reactions to human activities (beach)
– Age segmentation
– Etc
Assessing the impact of coastline and territortial
managment regulations
17. And next ?
• Delivery planned in june 2016 as a QGIS plugin
• Provide you a « Drone + Software » package
or All As A Service
• R&D improvments to manage turtles redundancy
• Involve public in turtles validation phase
• Other species detections are possible : rays,
sharks, dolphins, whales, dugongs…illegal
fishing boats
18. So what if
• You would become turtles angel’s eyes ?
• Goal: use it on 3 to 5 other sites until next ISTS
> The more we will have data, better it will work !
20. Any question ?
Gwénaël Duclos
Ecology and Vision Engineer
+33 (0)6 75 40 38 01
gwenael.duclos@wipsea.com
Editor's Notes
Hola, Good morning everybody, I’m Gwenael Duclos, I founded WIPSEA three years ago with the aim of using my skills in computer vision for ecology. We are now four people. I’m glad to be here to present you sea turtles semi-automatic mapping technics. They are still under development in a collaboration project we have with Mayotte Island MPA, French MPA Agency, Kélonia and Actimar. I would like to thank all the co-authors from their contributions in this project and this talk.
Why semi-automated mapping technics ? Because sea turtles need to be efficiently protected, we are all here for that cause ! In parallel with your big effort on nesting beaches protection and bycatch reduction, particular efforts are also made to protect foraging areas. The effectiveness of feeding habitat management and conservation plans requires abundance estimates of the population.
Aerial surveys appears to be the most appropriate technique to obtain these data thanks to their spatio-temporal coverage efficiency. But they cost a lot, are relatively dangerous due to low altitude and low speed required to make efficient observations.
Amanda Hodgson had the idea in 2010 to use drones to survey dugongs on the West Coast of Australia. I had the luck to work will her during my retraining internship. We built with two other researchers specialized in computer vison the first software capable of finding dugongs in photos taken by drone. You can see here that it was a big drone coming from Boeing developments for the US Army. We used it in Shark Bay to compare drone use with observers onboard aerial surveys
Katia Ballorain and Kelonia was following the dugong project and had the idea to use drones to map sea turtles in 2012. SEMMA-DRONE means Marine Megafauna Survey with DRONE. Reducing the costs of feeding habitats surveys to be able to increase their frequency is the main goal of this project. To do so, we aim at developping a semi-automated solution to speed up the overall process, from data acquisition by a drone aircraft to their representation in an Open Source GIS.
Data acquisition was done using Sensefly Ebee drone. It’s a totally autonomous flying wing capable to fly up to 45’ minutes without pilot. You have just to hand-launch it from the beach and retreive it as a freezebe or let it land on the beach. This drone is equipped with a RGB Canon camera with a 12Megapixels sensor and a focal length to reach a ground resolution of 3.5 cm/pixel. This camera is equipped with a picture stabilizer and a GPS. The camera shots are controlled by the drone auto-pilot so all is completely autonomous.
As I said this project is based in Mayotte island. This is a small french island in the Mozambique Channel near Madagascar.
The surveys a were made in N’Gouja reserve at the south west of the island.
So we covered less than a 1 kilometre square area with two flights at 50 m altitude. With a 40% overlay, those two flights produced around three hundred pictures. Note that this overlay will produce several redundant observations because a turtle can be on two or more pictures.
From GPS information and drone attitude angles recorded by the drone synchronously with pictures shots, ACTIMAR produced a projection model which give a geographic position to each pixel so we were able to map each picture on google map.
Here is a picture taken from the drone. Let’s do a quick test ! How many turtles do you see on this picture ? 1/2/3/4/5/6/7/8/9/10 ? There are 9 turtles.
So this little test give you the complexity of detecting and counting turtles on one picture, so imagine with 300 pictures for each flight…
This is the reason why SemmaDrone project includes the development of an algorithm, embedded in QGIS, to assist the user in image analysis phase by proposing estimated turtles’s positions in each image. This algorithm is in two steps: the first one automatically detect interesting candidates based on colour and shape features.
The second step is an object classification method based on a state of the art learning algorithm that was trained to recognize sea turtles. It gives a probability score to each candidate and you can see that turtles have a score above 70%, other objects such as rays or coral are lower but you can also see HERE that Nature entropy can create fake turtles with corals arrangements. This is the reason why we still give the user the final decision step.
Thanks to an accurate geo-referencing of the images, each turtle detected like so can be registered straighforward in the GIS layers to provide a spatio-temporal vision of the population evolution in the studied area. We are adding an interface to allow the user to validate each detection then add some informations such as the size, the activity and sometimes the sex when picture quality allow it.
At Mayotte MPA, these data are used to evaluate population size, foraging habitat use, Turtles reactions to human activities (beach), Age segmentation and so on.
This method will be used in the the future to assess the impact of coastline and territorial management regulations.
.
The official delivery of the first release is planned in june 2016 as a QGIS plugin.
We would like to test it in other sites: who is interested ? Duke University who already have such a drone ?
This method is not limited to drone use because the camera can be installed in a plane so we could test on a larger area.
And we are already working on using it on other species detection: rays, sharks, dolphins, whales, dugongs…and illegal boats
WIPSEA can provide you a drone and sofware package or do all the process as a service
The official delivery is planned in june 2016
We would like to test it in other sites: who is interested ? Duke University who already have such a drone ?
This method is not limited to drone use because the camera can be installed in a plane so we could test on a larger area.
And we are already working on using it on other species detection: rays, sharks, dolphins, whales, dugongs…as shown in our video Tuesday night.