|
Impact of jellyfish in human activities has been increasingly reported worldwide in recent years. Segments such
as tourism, water sports and leisure, fisheries and aquaculture are commonly damaged when facing blooms of
gelatinous zooplankton. Hence the prediction of the appearance and disappearance of jellyfish in our coasts, which
is not fully understood from its biological point of view, has been approached as a pattern recognition problem
in the paper presented herein, where a set of potential ecological cues was selected to test their usefulness for
prediction. Remote sensing data was used to describe environmental conditions that could support the occurrence
of jellyfish blooms with the aim of capturing physical-biological interactions: forcing, coastal morphology, food
availability, and water mass characteristics are some of the variables that seem to exert an effect on jellyfish
accumulation on the shoreline, under specific spatial and temporal windows.
A data-driven model based on computational intelligence techniques has been designed and implemented
to predict jellyfish events on the beach area as a function of environmental conditions. Data from 2009 over
the NW Mediterranean continental shelf have been used to train and test this prediction protocol. Standard
level 2 products are used from MODIS (NASA OceanColor) and MERIS (ESA - FRS data). The procedure for
designing the analysis system can be described as following. The aforementioned satellite data has been used as
feature set for the performance evaluation. Ground truth has been extracted from visual observations by human
agents on different beach sites along the Catalan area. After collecting the evaluation data set, the performance
between different computational intelligence approaches have been compared. The outperforming one in terms
of its generalization capability has been selected for prediction recall.
Different tests have been conducted in order to assess the prediction capability of the resulting system in
operational conditions. This includes taking into account several types of features with different distances in
both the spatial and temporal domains with respect to prediction time and site. Moreover the generalization
capability has been measured via cross-fold validation. The implementation and performance evaluation results
are detailed in the present communication together with the feature extraction from satellite data. To the best
of our knowledge the developed application constitutes the first implementation of an automate system for the
prediction of jellyfish appearance founded on remote sensing technologies.
|
