Pasar al contenido principal

Resultados relacionados

Resultados totales (Incluyendo duplicados): 1
Encontrada(s) 1 página(s)

DIGITAL.CSIC. Repositorio Institucional del CSIC
Publicación Artículo científico (JournalArticle)

Improving prediction region accuracy in marine animal movement with temporal fusion transformer

  • Medina Hernández, Jorge
  • Rodríguez-García, Jorge Pablo
  • McMahon, Clive R.
  • Sequeira, Ana M. M.
  • Eguíluz, Víctor M.
Predicting marine animal movements from satellite tracking data remains challenging, limiting conservation and ecosystem management efforts. To address this, we trained the Temporal Fusion Transformer (TFT) neural network on tracking data from 434 southern elephant seals to forecast locations and fill data gaps (imputation) within 7-day windows. Compared to state-space models, TFT reduced location errors by 15% and produced more efficient prediction regions, identifying where seals were likely to be found while using less area: a fivefold reduction for forecasting and 30-40% reduction for imputation. The model performed best near the continental shelf and at low-to-moderate movement speeds, with bathymetry, water temperature and current direction being the most influential environmental factors affecting the model output. When applied to new geographic regions not represented in the training dataset, model performance declined by approximately 30% across most evaluation metrics, indicating challenges in transferring learned patterns to unfamiliar environments. Our findings show that deep learning is a promising tool for analyzing large, sparse tracking datasets. The enhanced predictive capabilities have potential for dynamic conservation measures, such as forecasting the spatial evolution of animals to minimize conflicts with human activities and environmental disturbances., This research was funded by the Sustainable Ocean Alliance for the project Artificial Intelligence & Animal Movement (AIAM, ref. D017) project under the SOA Grants program. J.P.R. acknowledges funding from the Vicenç Mut program of Govern de les Illes Balears. This research was supported by Ministerio de Ciencia e Innovacion, Agencia Estatal de Investigación through the Maria de Maeztu Units of Excellences Program (project CEX2021-001164-M) funded by MICIU/AEI/10.13039/501100011033, and by project CSxAI (PID2024-157526NB-I00), funded by MICIU/AEI/10.13039/501100011033/FEDER, UE, and by an ARC DP210103091 (awarded to A.M.M.S. and V.M.E.). J.M.H. acknowledges financial support from MDM-2017-0711-19-1, funded by MICIU/AEI/10.13039/501100011033 and by FSE invierte en tu futuro. Data collection was supported by the French Polar Institute (program 109: PI. H. Weimerskirch and 1201: PI. C. Gilbert and C. Guinet), the SNO-MEMO and CNES-TOSCA, in collaboration with The Integrated Marine Observing System (IMOS) with data sourced from Australia’s Integrated Marine Observing System (IMOS) – IMOS is enabled by the National Collaborative Research Infrastructure Strategy (NCRIS)., With funding from the Spanish government through the "Maria de Maeztu Centre of Excellence" accreditation (CEX2021-001164-M)., Peer reviewed