Line of action:

Status:

Execution date:

General:

The Deep-RAMP – Deep-learning proposal to improve the management of the Network of Marine Protected Areas of the North Atlantic Demarcation aims to cover the needs of automation in the identification and inventory of the structuring species of the vulnerable benthic habitats of the Marine Natura 2000 Network, and will allow the monitoring of the effects that the management measures of these areas may cause on benthic habitats of special vulnerability.

Specific:

1. To advance in the technological development that makes it possible to monitor funds through image analysis techniques based on artificial intelligence algorithms.

2. Automation in the identification and inventory of the structuring species of the vulnerable benthic habitats of the Natura 2000 Network.

3. Improve the monitoring of marine RN2000 areas, automating the analysis of information to assess the effects that the management measures of these areas may have on benthic habitats of special vulnerability.

A0 Project coordination and monitoring tasks

A1. Target species selection

A2. Data collection from all zones

2.1. Collection of images and information – El Cachucho

2.2. Collection of images and information – Avilés

A3. Algorithm testing

A4. Training of the neural network or Deep-learning algorithm

4.1. Collection of catches of target species

4.2. Neural Network Training

A5. Implementation of a customized system for RN2000 Marina

A6. Validation of results

A7. Outreach activities

7.1. Press dissemination and results dissemination day

7.2. Scientific publications at specialised conferences

7.3. Scientific dissemination to society and the general public

During the development of the project, a series of algorithms have been obtained that allow the automatic analysis of images of deep bottoms in Natura 2000 Network areas. These tools, based on artificial intelligence or deep learning algorithms, manage to mark structuring species automatically and in a very short time, species to be protected by means of figures within the framework of the Natura Network.

The project is a first step in the implementation of nets and algorithms tested in oceanographic campaigns, which can optimize efforts, allowing the identification and classification of benthic species of interest and that can be used as bioindicators, contributing to the study of the benthic systems that make up the NATURA 2000 NETWORK, a key aspect to manage Marine Protected Areas.

For the execution of the project, structuring species of vulnerable habitats were selected in the area of the Avilés Submarine Canyon System, considering those species that, due to their size or morphological characteristics, allow their identification, based exclusively on images:

• A cnidarian: Dendrophyllia cornígera
• Two sponge morphotypes: Artemisina transiens and Phakellia ventilabrum
• Three stars: Ophiothrix fragilis, Ophiura ophiura and sp. III
• A feather star: Leptometra celtica

In the cold-water coral reef area of the Avilés Submarine Canyon System, the following structuring species have been selected:

• Madrepora oculata
• Desmophyllum pertusum
• Lepiophates sp.
• Phoenix watering station
• Cerianthus lloydi

Likewise, information from previous oceanographic campaigns was collected and photographs and video frames were selected for the application and validation of the image analysis algorithms developed. The data used come from oceanographic campaigns carried out within the framework of the LIFE-IP-INTEMARES project “Integrated, innovative and participatory management of the Natura 2000 Network in the Spanish marine environment”. In the same way, all the images used have been acquired through the operation of the remotely operated towed underwater vehicle (ROTV Politolana), using the video camera and different on-board cameras. The auxiliary data, such as the precise positioning of the images, have been obtained in the same way, with the ROTV system, and the acoustic positioning system for underwater vehicles carried by the ships.

In this way, during the development of the project, the algorithms or neural networks selected for automatic recognition were trained and tested. To do this, the images labeled by an expert manually and the determination of the weights that acted on the different detection layers were used. In the validation of the results of the rocky circalitoral zone, the accuracy of the model (correct classification at the species level among all the objects in the network), the recall (level of detection of specimens within a species) and the F1-score (indicator that brings together both quality parameters in a single number) were taken into account. These metrics were validated against one of the transects with the highest abundance in species (transect IA418_TF_23), which consists of 188 images. The optimal threshold was set at 0.75 for the metrics, obtaining for the species of the genera Phakellias, Dendrophyllias and Artemsinas a mean F1-score of 0.75, a mean recall of 0.72 and an accuracy of 0.78.

In the cold-water reef area, the accuracy value of the model was taken into account, assessing whether the pixel had been correctly labeled, after scanning the entire image, comparing it with the labeling carried out by the expert. The metrics were validated against a set of 11 high-resolution images, previously labeled by an expert, as well as against a series of frames extracted from a video-transect. An accuracy of 89.4% for the set of the images and 88.9% for the video was obtained.

The project included dissemination activities, with a presence at international conferences and scientific journals to publicize the progress made.