Baseline cruise before deep sea mining
Onboard the Kamei research vessel, Dhugal Lindsay (Japan) surveyed a prospective deep sea mining site. This cruise helps establish a baseline of the ecosystem in case mining is approved, in order to assess its potential impact. This is one of the "demonstrators" of the WWWPIC project, in which the image handling infrastructure is used to process data and exploit it to answer an ecological question.
During this cruise, a Remotely Operated Vehicle was deployed with a stereo camera and a Video Plankton Recorder was attached to the CTD profiler (pictured).
EcoTaxa's API is live (and documented)
Since the start of September, Laurent Salinas (France), the developper supervising the improvement in EcoTaxa through WWWPIC, has tackled the task to move all the functionality to a backend server and expose them through an Application Programming Interface (API). This API will allow machine-to-machine interoperability between EcoTaxa and other databases.
The functions to move are numerous and their interface not always easy to specificy, but the API is now live and self-documented. It is used by third parties already. Now it will steadily grow with time, as more functions are moved there.
Efficient representation of plankton visual appearance
For plankton classification by computers to work efficiently, relevant features must be extracted from the images. These features are a summary of the visual aspect of the image. Convolutional Neural Networks (CNNs) are particularly effective at image classification because, during model training, the features are optimised to improve classification performance. One way to assess the quality of those features is to represent them in a reduced space, through a t-SNE for example, and check that categories (in color in the figure on the right) are well separated. This is what Antonio Goulart and Alexandre Morimitsu (Brazil) did and indeed, the trained network led to a clear separation of the different plankton taxa. They presented a pilot study at the 2021 OCEANS Conference. Now this model and features can be used for further endeavors.
Images from the deep Coral Sea
Dhugal Lindsay (Japan), recently took part in the Schmidt Ocean Institute-sponsored cruise “Seamounts, canyons and reefs of the Coral Sea” via telepresence in order to collect images of gelatinous macroplankton for the wwwPIC project. In the coming months, taxonomically-vetted images of various species will be harvested from the original high quality video footage to help train a machine learning model to recognize the species in Remotely-Operated Vehicle (ROV) footage. The dives, which were live-streamed on YouTube, can be found, with Lindsay’s commentary, at the following links:
ROV Dive 373 - Coral Sea Marine Park - Cairns Seamount (Streamed live on 6 Aug 2020)
ROV Dive 376 - Coral Sea Marine Park - Herald Cays (8 Aug 2020)
ROV Dive 377 - Coral Sea Marine Park - Malay Reef (10 Aug 2020)
ROV Dive 389 - Coral Sea Marine Park - Osprey Reef (21 Aug 2020)
ROV Dive 392 - Coral Sea Marine Park - Tregrosse Reefs (25 Aug 2020)
New imaging sensor integrated on a glider
Marc Picheral, engineer in Villefranche (France), supervised the integration of the UVP6 imaging sensor that he invented on the SeaExplorer glider from ALSEAMAR. The glider carried out a full back and forth transect across the Ligurian current and the UVP6 worked nominally. This test deployment revealed hiccups with other sensors (that is what tests are for afterall!) but, on the plankton imaging side, we are ready to deploy this platform next spring. The glider equipped with the UVP6 will generate a large number of images which will be processed and integrated into EcoTaxa. This data will serve as a demonstration of the use of the infrastructure in the last years of the project, as part of Thelma Panaïotis' PhD thesis.
EMODnet Biology Workshop
Laurent Salinas (France), the developer working on EcoTaxa for WWWPIC, followed the EMODnet Biology (all online...) workshop, to learn more about how to send data from EcoTaxa to EMODnet and then OBIS. After these two weeks, a first implementation of an export format for EMODnet is ready and passes the first quality tests. Now we "just" need to make it bullet proof ;-)
A comprehensive learning set for the ISIIS instrument
Moritz Schmid (USA), who works as a post doc on the project, prepared an extensive data of ISIIS images from the Northern California Current and uploaded it on EcoTaxa. With ~83,000 images distributed alsmost evently in 170 classes, it will allow to test various classification techniques and bootstrap the creation of other datasets for the same instrument.
IFB will host EcoTaxa-EU
The French Institute of Bioinformatics is a national infrastructure dedicated to storing and computing with biological data, primarily nucleic acid sequences. Quantitative images, that are the target of WWWPIC, are small and numerous and the technical requirements to store and process them are close to those for nucleic acids sequences.
After examining the requirements, IFB has agreed to host an instance of EcoTaxa to serve as the European node of the WWWPIC infrastructure. In the meantime, the code is being adapted for running on such professional hosting services, by modularising it and packaging it into Docker containers. This will help its deployment in other countries.
EcoTaxa in Chinese
As part of the wider collaboration network around WWWPIC, our collaborators from the company Watertools in China have deployed an instance of EcoTaxa and translated most of its interface in Chinese. We will continue to work towards easing such internationalisation by separating the code and the messages in the user interface, but it already looks very cool with Chinese characters!
We will also work with Watertools and the Institute of Oceanology of Chinese Academy of Sciences to find a good hosting place for this instance and integrate it in the network of EcoTaxa nodes that WWWPIC is building.
Intern Adeb Zaoui (France) started working remotely because of the current sanitary crisis. Under the supervision of Éric Debreuve and Cédric Dubois, he has explored a hierarchical image classification deep neural network on a standard dataset. He was able to reproduce the results of a recent paper on the topic. Now, he will be moving to plankton images.
UniEuk + WoRMS
The UniEuk project provides the current taxonomic backbone of the EcoTaxa application, which is one of the the main tools in the WWWPIC project. Having a reliable, universal taxonomy is the starting point for any ecological dataset.
To be able to provide such datasets to other databases, they need to share a common taxonomy. Uploading datasets to OBIS is one important objective of WWWPIC and OBIS uses the well-known WoRMS taxonomy. UniEuk has recently integrated the WoRMS taxonomy, to complete its coverage and so that individual taxa identifiers can be translated from one convention to the other. This is an important first step towards sending data to OBIS!
Reference image datasets
One goal of WWWPIC is to provide reference datasets of curated images to train new machine learning models. We started with a dataset of 72k images for the ISIIS instrument. It is currently being complemented by two other datasets for the same instrument but where each organism has been carefully segmented manually and a larger, more diverse dataset from the Northern California Current being prepared now.
Beyond ISIIS, we improved upon a reference dataset for the ZooScan instrument, that currently holds 1.4M images sorted in ~160 classes, and are currently building a new one for the FlowCam instrument.