Modelagem da cobertura de gelo marinho nos mares Antárticos de Weddell, Bellingshausen e Amundsen com uso de redes neurais artificiais
Sea ice, covering approximately 7% of the surface of the Earth’s oceans, is a fundamental climatic component for studying climate in the polar regions, mainly because it acts as a natural barrier at the ocean-atmosphere interface that restricts the exchange of heat, mass, and momentum between the se...
Na minha lista:
Autor principal: | |
---|---|
Outros Autores: | |
Formato: | doctoralThesis |
Idioma: | pt_BR |
Publicado em: |
Universidade Federal do Rio Grande do Norte
|
Assuntos: | |
Endereço do item: | https://repositorio.ufrn.br/handle/123456789/45702 |
Tags: |
Adicionar Tag
Sem tags, seja o primeiro a adicionar uma tag!
|
Resumo: | Sea ice, covering approximately 7% of the surface of the Earth’s oceans, is a fundamental
climatic component for studying climate in the polar regions, mainly because it acts as
a natural barrier at the ocean-atmosphere interface that restricts the exchange of heat,
mass, and momentum between the sea and the air, and reflects much of the incident
solar radiation. Satellite observations since the 1970s indicate an Arctic with increasingly
thinner and younger sea ice, accompanied by a decline in its extent. While some areas
of Antarctica (e.g. Ross Sea and Weddell Sea) have shown a slight increase in the extent
of this climate component for the same period. In this context, this study had as main
objective to evaluate the potential predictability of sea ice cover with the application of
ANNs techniques (MLP, LSTM and CNN-LSTM) in 3 Antarctic seas, they are Weddell,
Bellingshausen and Amundsen. The data used were the monthly reanalyses from Era-5,
for the period 1979 to 2019. SARIMAX models served as reference values for gauging the
accuracy of the predictions with ANNs. Analyzing the errors (RMSE and MAE) of Sea
Ice Concentration (SIC) for all the studied seas it was observed that the CNN-LSTM was
exceeded only in the months with values lower than ±25%, and the maximum differences
between the errors in these events did not reach values above 5% ,as what occurred in
the months of July, September and October in the Weddell Sea. With these results it was
possible to infer that the CNN-LSTM model was the one that most accurately predicted
the periods with the greatest differences (> 0.15) of SIC. With the analyses of the spatial
distributions of the SIC differences it can be affirmed that the delimitations of the fields
with differences greater than 0.15 were also better predicted with the coupling of the CNN
and LSTM architectures, as occurred with the negative anomaly of June in the Weddell
Sea (MAE < 0.15) and the positive one, in September, in the Amundsen Sea (MAE < 0.10). |
---|