A tropical Atlantic dynamics analysis by combining machine learning and satellite data

Arnault, Sabine; Thiria, Sylvie; Crépon, Michel; Kaly, François

doi:10.1016/j.asr.2020.09.044

Cited by 9 publications

(8 citation statements)

References 57 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…The SOM analysis has been applied in many fields (Kaski et al, 1998;Oja et al, 2003;Hong et al, 2004;Hong et al, 2005). In oceanography, it was found useful for the extraction and analysis of currents, wind, SST, Sea Surface Height and Sea Surface Salinity patterns (Richardson et al, 2003;Liu and Weisberg, 2005;Liu et al, 2006a;Mihanovićet al, 2011;Jouini et al, 2016;Liu et al, 2016;Arnault et al, 2021).…”

Section: Concept and Architecturementioning

confidence: 99%

“…We use the MATLAB SOM-Toolbox version 2.0 (Vesanto et al, 2000;Kohonen, 2014), which is an efficient SOM implementation developed by the Laboratory of Information and Computer Science in the Helsinki University of Technology freely available at http://www.cis.hut.fi/projects/somtoolbox/. This toolbox was successfully used previously in oceanographic applications (Jouini et al, 2016;Liu et al, 2016;Liu et al, 2006a;Liu et al, 2006b;Arnault et al, 2021). The mapping quality of the SOM can be quantified through the average quantization error (QE) and topographic error (TE) which are returned by the toolbox (Liu et al, 2006a).…”

Section: Concept and Architecturementioning

confidence: 99%

“…With global coverage and long-term biological and physical observations, satellite remote sensing is an effective technology for understanding ocean processes. Unsupervised ML is a major class of ML that can cluster and reveal significant patterns in high dimensional datasets such as satellite ocean observations (Arnault et al, 2021;Sonnewald et al, 2021). Unsupervised ML identifies relationships among satellite observations of environmental parameters and reduces their large dimensionality, while removing unstructured noise (Boukabara et al, 2021;Rolf et al, 2021;Sonnewald et al, 2021).…”

Section: Introductionmentioning

confidence: 99%

“…For example, the application of kmeans, a space-partitioning algorithm, to satellite ocean colour data identified regions with distinct marine biological activity (Ardyna et al, 2017). The SOM, a neural network-based clustering algorithm, found several applications in extracting spatial and temporal characteristic ocean current patterns using satellite sea surface height, current and surface salinity data (e.g., Liu et al, 2016;Arnault et al, 2021). A GAN, a method that generates new samples with similar distributions as the learning dataset, was used to reconstruct sea surface heights of the North Sea (Zhang et al, 2020;Yu and Ma, 2021).…”

Section: Introductionmentioning

confidence: 99%

“…The high variability of the AC and the lack of continuous and repeated in-situ observations, make satellite data and the SOM algorithm well suited to our study goal. The SOM algorithm (Kohnen, 1982), which has seen a wider application in recent years in ocean science (Liu and Weisberg, 2011;Arnault et al, 2021), has the advantage of higher discriminant power (distinguishing between the different clusters) and the ability to deal with the nonlinear aspects of the biophysical phenomena we aim to analyze, as compared to other ML (e.g., K-means; Badran et al, 2005) or statistical (e.g., the Empirical Orthogonal Functions [EOF]; Lorenz, 1956) methods (Jouini et al, 2016).…”

Section: Introductionmentioning

confidence: 99%

See 4 more Smart Citations

Earth Observation and Machine Learning Reveal the Dynamics of Productive Upwelling Regimes on the Agulhas Bank

et al. 2022

View full text Add to dashboard Cite

The combined application of machine learning and satellite observations offers a new way for analysing complex ocean biological and physical processes. Here, an unsupervised machine learning approach, Self Organizing Maps (SOM), is applied to discover links between surface current variability and phytoplankton productivity during seasonal upwelling over the Agulhas Bank (South Africa), from 23 years (November-March 1997-2020) of daily satellite observations (surface current, sea surface temperature, chlorophyll-a). The SOM patterns extracted over this dynamically complex region, which is dominated by the Agulhas Current (AC), revealed 4 topologies/modes of the AC system. An AC flowing southwestward along the shelf edge is the dominant mode. An AC with a cyclonic meander near shelf is the second most frequent mode. An AC with a cyclonic meander off shelf and AC early retroflection modes are the least frequent. These AC topologies influence the circulation and the phytoplankton productivity on the shelf. Strong (weak) seasonal upwelling is seen in the AC early retroflection, the AC with a cyclonic meander near shelf modes and in part of the AC along the shelf edge mode (the AC with a cyclonic meander off shelf mode and in part the AC along the shelf edge mode). The more productive patterns are generally associated with a strong southwestward flow over the central bank caused by the AC intrusion to the east Bank or via an anticyclonic meander. The less productive situations can be related to a weaker southwest flow over the central bank, strong northeast flow on the eastern bank, and/or to a stronger northwest flow on the central bank. The SOM patterns show marked year-to-year variability. The high/low productivity events seem to be linked to the occurrence of extreme phases in climate variability modes (El Niño Southern Oscillation, Indian Ocean Dipole).

show abstract