Prediction of missing values and detection of ‘exceptional events’ in a chronological planktonic series: a single algorithm

Abstract

The detection of extreme events is of primary importance because they often change the initial conditions of a dynamic system. However, the definition of what constitutes an extreme or exceptional event is unclear; what threshold or which rate of occurrence delineates an anomaly? An alternate and precisely specified type of definition might be an event which cannot be predicted by a particular model at a chosen probability. Missing values are unfortunately characteristic of biological oceanographic time series. This characteristic precludes a great deal of numerical treatments. Consequently, several interpolation techniques have been proposed to predict missing values. Most of them are not adequate for planktonic data which are characterized by high heterogeneity. An iterative approach based on the principles of the eigenvectors filtering (EVF) method is examined. The limits of the technique are determined through simulation. The same method is then applied for the detection and definition of extreme events. We first apply a crude method to select some maximal or minimal values in a data series (the extreme events), then the selected values are coded as missing values, and finally we evaluate how well the EVF is able to reproduce the original ‘extreme’ values. These simulations provide insight into why large peaks (or holes) can be identified as extremes events or not, based on the degree of their prediction.