Seguimiento del estado de salud de los ecosistemas mediante la deteccion automatica de cantos de aves usando deep learning
Seguimiento del estado de salud de los ecosistemas mediante la deteccion automatica de cantos de aves usando deep learning
Principal investigator
Irene Mendoza
Financial institution
MIN CIENCIA E INNOVACION
Fecha de inicio
Fecha de fin
Code
TED2021-129871A-I00
Department
Ecology and Evolution
Researchers
Jordano, Pedro
Brief description
Estamos viviendo actualmente una crisis ecológica en la que las especies, las interacciones entre ellas y los servicios que la naturaleza brinda al ser humano se están perdiendo a un ritmo sin precedentes. Por lo tanto, es urgente desarrollar sistemas de diagnóstico de la salud de los ecosistemas que sean rápidos, confiables, replicables y automáticos. Los cambios en la migración y abundancia de las especies de aves cantoras son indicadores del estado de salud de los ecosistemas, ya que las fechas de llegada y salida de las especies de aves se ven afectadas por el cambio climático. El seguimiento de la diversidad de aves se ha realizado hasta ahora mediante censos de expertos, pero los actuales avances tecnológicos nos permiten ampliar increíblemente las escalas espaciales y temporales de estudio gracias al seguimiento acústico pasivo. El principal desafío de éste es que se generan rápidamente petabytes de datos, lo cual excede lo que un experto humano puede anotar manualmente en un tiempo razonable. Por tanto, se impone no solo registrar de forma automática el canto de las aves, sino también la detección de las mismas. Esta propuesta tiene como objetivo hacer un seguimiento automático de la diversidad de aves cantoras desarrollando las herramientas bioinformáticas y de deep learning necesarias para comprender los cambios espacio-temporales en las comunidades de aves, con el fin de generar predicciones precisas en escenarios futuros. Para ello, estableceremos una ciberinfraestructura de seguimiento del canto de las aves en el Parque Nacional de Doñana mediante grabadoras remotas de código abierto combinadas con procesadores Raspberry Pi, aprovechando la infraestructura científico-técnica singular ya existente en Doñana (ICTS). Pretendemos automatizar la identificación de las especies mediante redes neuronales convolucionales. Esta propuesta multidisciplinar combinará técnicas tanto ecológicas como de ciencia de datos para resolver tres tareas específicas: 1) Evaluar el efecto del cambio climático en las comunidades de aves de Doñana; 2) Automatizar el proceso computacional de identificación de especies de aves en grandes conjuntos de datos de audio 3) Pronosticar cambios futuros de las comunidades de aves según diferentes escenarios de cambio climático. Los investigadores involucrados en esta propuesta tienen experiencia tanto en ecología como en ciencia de datos y aplicarán su conocimiento profundo de la comunidad de aves de Doñana a las últimas técnicas de deep learning aplicadas al reconocimiento de audio. Esta propuesta tiene un doble impacto: por un lado, nos permitirá conocer de forma fidedigna los cambios en la avifauna como forma de conocer el estado de salud del ecosistema de Doñana; por otro, permitirá un enorme desarrollo de las técnicas de seguimiento automático de la biodiversidad, allanando el camino para establecer una red de seguimiento automático a escala nacional o europea // We are living in an ongoing ecological crisis in which species, interactions among them and services provided by nature to humans are
being lost with an unprecedented pace. It is therefore urgent to develop diagnosis systems of the health of ecosystems that are fast, reliable, replicable, and automatic. Changes in the migration times and abundance of birdsong species are indicators of the health status of ecosystems, as the arrival and departure dates of avian species are changing as a consequence of climate change. Monitoring avian diversity has been performed so far with the use of expert surveys, but technological advances allow us now to incredibly enlarge spatial and temporal scales thanks to passive acoustic monitoring. The main challenge of automatically monitoring birdsong acoustic diversity using recorders is that petabytes of data are very rapidly generated, which exceeds what a human expert can manually annotate in reasonable time. Therefore, it is imposed not only to automatize bird recordings, but also bird detection, which is now a feasible task thanks to the advance of machine learning techniques. This proposal aims to monitor birdsong diversity in an automated manner developing the bioinformatics and deep-learning approaches necessary for understanding spatio-temporal changes in bird communities, in order to generate accurate predictions under future scenarios. For this, we will establish a birdsong monitoring cyberinfrastructure at Doñana National Park using remote open-source recorders combined with Raspberry Pi processors and automatize the identification of the species using deep learning techniques, taking advantage of the large infrastructure already existent at Doñana (ICTS). This is an intrinsically multidisciplinary proposal that will combine both ecological and data science techniques to solve three specific tasks 1) Analyse climate-change effects on avian communities of Doñana; 2) Automatize the computational process of species identification in large datasets of birdsong diversity monitoring 3) Forecast changes in bird communities under future scenario of climate change. Researchers involved in this proposal have both an ecological and data-science background and will apply their expert knowledge of the Doñana bird community to the latest techniques of deep learning applied to audio recognition. The impacts derived from this proposal are two-sided: first, it will allow us to extract trustable information of changes in birdsong fauna as proxies of the health status of the ecosystems, and also, it will largely increase the power of new technologies and deep-learning processes applied to biodiversity monitoring, paving the road for more extensive national or European automatic monitoring surveys.
being lost with an unprecedented pace. It is therefore urgent to develop diagnosis systems of the health of ecosystems that are fast, reliable, replicable, and automatic. Changes in the migration times and abundance of birdsong species are indicators of the health status of ecosystems, as the arrival and departure dates of avian species are changing as a consequence of climate change. Monitoring avian diversity has been performed so far with the use of expert surveys, but technological advances allow us now to incredibly enlarge spatial and temporal scales thanks to passive acoustic monitoring. The main challenge of automatically monitoring birdsong acoustic diversity using recorders is that petabytes of data are very rapidly generated, which exceeds what a human expert can manually annotate in reasonable time. Therefore, it is imposed not only to automatize bird recordings, but also bird detection, which is now a feasible task thanks to the advance of machine learning techniques. This proposal aims to monitor birdsong diversity in an automated manner developing the bioinformatics and deep-learning approaches necessary for understanding spatio-temporal changes in bird communities, in order to generate accurate predictions under future scenarios. For this, we will establish a birdsong monitoring cyberinfrastructure at Doñana National Park using remote open-source recorders combined with Raspberry Pi processors and automatize the identification of the species using deep learning techniques, taking advantage of the large infrastructure already existent at Doñana (ICTS). This is an intrinsically multidisciplinary proposal that will combine both ecological and data science techniques to solve three specific tasks 1) Analyse climate-change effects on avian communities of Doñana; 2) Automatize the computational process of species identification in large datasets of birdsong diversity monitoring 3) Forecast changes in bird communities under future scenario of climate change. Researchers involved in this proposal have both an ecological and data-science background and will apply their expert knowledge of the Doñana bird community to the latest techniques of deep learning applied to audio recognition. The impacts derived from this proposal are two-sided: first, it will allow us to extract trustable information of changes in birdsong fauna as proxies of the health status of the ecosystems, and also, it will largely increase the power of new technologies and deep-learning processes applied to biodiversity monitoring, paving the road for more extensive national or European automatic monitoring surveys.