Research output: Contribution to conference › Abstract › Scientific

- EGU2018-15552-0
36 KB, PDF-document

Traditionally, quantifying climate change induced uncertainty in ecological indicators requires stochastic simulation

with a chain of physically-based models describing various processes such as hydrodynamics, waves,

sediment transport and ecology. Such Monte Carlo based simulation on the entire model chain, especially with

large sample size, is however computationally expensive and often unfeasible.

In this paper, it was investigated how regression models can potentially replace physically-based models

and predict chlorophyll-a concentration directly from meteorological variables. Since several correlated meteorological

variables are used to estimate one ecological response variable, and thus a multi-collinearity problem is

present, Partial Least Squares (PLS) regression is considered to be a favourable supervised technique. On the other

hand, the climate change projection dataset at hand is multidimensional. This is due to the fact that it contains

several variables which are not only varying over time but also over space (spatially distributed). Consequently,

a multiway regression model should be applied which can account for the spatial dimension. The multiway

PLS regression (N-PLS) algorithm is a promising candidate for this purpose. The N-PLS is an extension of the

ordinary two-way PLS regression algorithm to multi-way data, where essentially the bilinear model of predictors

is replaced with a multilinear model.

In order to test its efficiency, the N-PLS algorithm was compared with other unsupervised and supervised,

two-way and multi-way techniques using both synthetic and real datasets. The latter dataset consists of meteorological

variables from KNMI (Royal Netherlands Meteorological Institute) and chlorophyll-a concentrations

obtained from the Delft3D WAQ ecological model. Firstly, it was confirmed that supervised techniques should

be favoured over unsupervised ones, due to their ability to include correlation to the response variable which

reduces prediction error. Moreover, the results suggest that by applying multi-way methods improvements can

be achieved in the prediction accuracy. The magnitude of these improvements is, however, case dependent. In

conclusion, it was found that N-PLS, as a supervised multi-way method, is a promising regression model for

the above mentioned purpose. Finally, due to the fast simulation time of the algorithm, it could be suitable for

stochastic simulation with large sample size for the assessment of climate change induced uncertainty in coastal

ecosystem indicators. Future work will focus on applying the fitted N-PLS model to EURO-CORDEX climate

change projections and quantify related uncertainties in the Wadden Sea ecosystem.

with a chain of physically-based models describing various processes such as hydrodynamics, waves,

sediment transport and ecology. Such Monte Carlo based simulation on the entire model chain, especially with

large sample size, is however computationally expensive and often unfeasible.

In this paper, it was investigated how regression models can potentially replace physically-based models

and predict chlorophyll-a concentration directly from meteorological variables. Since several correlated meteorological

variables are used to estimate one ecological response variable, and thus a multi-collinearity problem is

present, Partial Least Squares (PLS) regression is considered to be a favourable supervised technique. On the other

hand, the climate change projection dataset at hand is multidimensional. This is due to the fact that it contains

several variables which are not only varying over time but also over space (spatially distributed). Consequently,

a multiway regression model should be applied which can account for the spatial dimension. The multiway

PLS regression (N-PLS) algorithm is a promising candidate for this purpose. The N-PLS is an extension of the

ordinary two-way PLS regression algorithm to multi-way data, where essentially the bilinear model of predictors

is replaced with a multilinear model.

In order to test its efficiency, the N-PLS algorithm was compared with other unsupervised and supervised,

two-way and multi-way techniques using both synthetic and real datasets. The latter dataset consists of meteorological

variables from KNMI (Royal Netherlands Meteorological Institute) and chlorophyll-a concentrations

obtained from the Delft3D WAQ ecological model. Firstly, it was confirmed that supervised techniques should

be favoured over unsupervised ones, due to their ability to include correlation to the response variable which

reduces prediction error. Moreover, the results suggest that by applying multi-way methods improvements can

be achieved in the prediction accuracy. The magnitude of these improvements is, however, case dependent. In

conclusion, it was found that N-PLS, as a supervised multi-way method, is a promising regression model for

the above mentioned purpose. Finally, due to the fast simulation time of the algorithm, it could be suitable for

stochastic simulation with large sample size for the assessment of climate change induced uncertainty in coastal

ecosystem indicators. Future work will focus on applying the fitted N-PLS model to EURO-CORDEX climate

change projections and quantify related uncertainties in the Wadden Sea ecosystem.

Original language | English |
---|---|

Publication status | Published - 2018 |

Event | EGU General Assembly 2018 - Vienna, Austria Duration: 8 Apr 2018 → 13 Apr 2018 https://www.egu2018.eu/ |

Conference | EGU General Assembly 2018 |
---|---|

Abbreviated title | EGU 2018 |

Country | Austria |

City | Vienna |

Period | 8/04/18 → 13/04/18 |

Internet address |

ID: 45688094