Elsevier

Neurocomputing

Volume 30, Issues 1–4, January 2000, Pages 65-70
Neurocomputing

Neural network modelling for environmental prediction

https://doi.org/10.1016/S0925-2312(99)00144-7Get rights and content

Abstract

We describe the choice and assessment of neural network and statistical methods for data modelling, feature selection and forecasting. We deal in particular with how empirical environmental and Earth observation data can be used in conjunction with physical simulation models.

Introduction

One theme in the environment and climate Neurosat (“Processing of Environmental Observing Satellite Data with Neural Networks”) project relates to prediction of oceanic upwelling off the Mauretanian coast, using sea surface temperature (SST) images, and real and model meteorological data for the year 1982. Upwelling is the periodic replenishment of coastal surface waters with cold deep water, which has various attendant effects.

The data available to us mainly consist of the following.

The overall themes of our work are data and information fusion, empirical and model-based; handling data which is characterized by many uncertainties and numerous missing cases; and the development of “data-driven” pattern recognition and neural network methods. We are seeking an answer to the question as to whether such methods are an alternative to, or are complementary to, to large physical simulation and modeling systems.

Data exploration and selection was discussed in [2], [5]. Some aspects of this initial phase of the work were as follows. All empirical (observed) SST data were used, but we immediately faced the problem of missing values on a massive scale (over 70% of pixel values were missing, due directly or indirectly to cloud cover). Meteorological and environmental dynamics are mostly local, which led to the decision to cater for such local behaviour by nonlinear and locally piecewise methods. Hence we used a clusterwise regression method for imputation and nowcasting. Latency period of wind forcing of 10 or 11 d was assumed, which established the dimensionality of the problem. An approach to imputation of missing SST values was developed using clusterwise regression on wind/radiation, and spatial and temporal interpolation.

Section snippets

Learning from the Ispramix oceanographic model

Independently of the data imputation, another investigation was started which was based on ocean (physical) model output. The Ispramix model can provide complete output data, but at the cost of extensive computation time, and discrepancy from observed data. The latter difficulty can be mitigated by assimilation into the model of empirical data. The work reported below does not use assimilation, so that model output data approximates the reality. Use of assimilation is planned for the future.

If

Conclusion

Further details on the work described here, including animations and background information, can be found at http://hawk.infm.ulst.ac.uk:1998/neurosat (or by emailing the authors, [email protected]).

We have explored both neural network modelling, and the coupling of such data processing with (i)  empirically observed data, and (ii)  model output of the physical dynamics. Our data-driven methodology therefore takes account of the physics of the domain studied.

We have noted two directions for

Acknowledgments

Particular acknowledgment is due to W. Eifler, M. Ouberdous and E. Demirov, Joint Research Centre, SAI – Marine Environment Unit Data Assimilation sector, TP69, Ispra, Italy, for data, domain knowledge and many discussions. We also thank M. Crépon, S. Thiria and other consortium partners for feedback and comments received in discussions and presentations of this work.

References (5)

  • E. Demirov, W. Eifler, M. Ouberdous, N. Hibma, A 3-D finite volume free surface model for ocean simulations on parallel...
  • F. Murtagh, G. Zheng, J. Campbell, A. Aussem, M. Ouberdous, E. Demirov, W. Eifler, M. Crépon, in: R. Payne, P. Green...
There are more references available in the full text version of this article.
View full text