Por favor, use este identificador para citar o enlazar a este item: http://hdl.handle.net/10261/260696
COMPARTIR / EXPORTAR:
logo share SHARE BASE
Visualizar otros formatos: MARC | Dublin Core | RDF | ORE | MODS | METS | DIDL | DATACITE

Invitar a revisión por pares abierta
Título

Artificial Inteligence Techniques for Real-Time Monitoring of Fishing Catches: The iObserver

AutorAntelo, L. T. CSIC ORCID; Ovalle, Juan Carlos CSIC; Vilas Fernández, Carlos CSIC ORCID; Pérez Martín, Ricardo Isaac CSIC ORCID
Fecha de publicación2021
CitaciónWorld Fisheries Congress (2021)
ResumenSustainability is a basic premise for the economic and social future of fisheries and the main objective of fishing policies, such as the Common Fisheries Policy of the European Union. An immense challenge faced by sustainable fisheries management and policies is that of finding cost-effective monitoring methods. Several remote electronic monitoring devices (REM) systems are available to that purpose; however, many of them exhibit a number of drawbacks that prevent its generalization among fleets (e.g. off-line evaluation of catches by the so-called dry observers in land, crew interferences and distrust, etc.). To overcome these drawbacks, robust and reliable innovative technologies for registering captures are required. In this regard, we have developed the iObserver, an electronic device for automatic identification and quantification of the whole catch on board fishing vessels. The iObserver is installed in the fishing park, over the conveyor belt, just before the fishing separation zone. The system takes images of everything that crosses this conveyor belt during the separation process. The recognition software automatically analyzes every image, identifies all the individuals, estimates their length and generates a report containing the results. The Deep Learning algorithms for species recognition algorithms included in the iObserver use Convolutional Neural Networks. Two models were created with transfer learning and data augmentation techniques: an instance segmentation model for specimen detection and classification and a regression model for fish length estimation. To train and test both models, a set of about 6,000 images were labeled with mask, species and length of each individual for a selection of 14 of the most relevant species for the local fisheries. Preliminary results were promising: precision and recall of 98% and 95% respectively for species recognition and 3.2% mean absolute percentage error for length estimation in the test data set. Species identification and quantification results are combined with data supplied by the vessel instrumentation, such as position, course, velocity. Finally, these georeferenced total catch data (a small CSV file) is sent, in real time, to an onshore center for further displaying and analysis enabling their use, either by ship-owners or policy makers, for efficient fishing activity management purposes
DescripciónWorld Fisheries Congress 2021, 20-24 September, Adelaide (Australia)
URIhttp://hdl.handle.net/10261/260696
Aparece en las colecciones: (IIM) Comunicaciones congresos




Ficheros en este ítem:
Fichero Descripción Tamaño Formato
accesoRestringido.pdf15,38 kBAdobe PDFVista previa
Visualizar/Abrir
Mostrar el registro completo

CORE Recommender

Page view(s)

74
checked on 27-mar-2024

Download(s)

9
checked on 27-mar-2024

Google ScholarTM

Check


NOTA: Los ítems de Digital.CSIC están protegidos por copyright, con todos los derechos reservados, a menos que se indique lo contrario.