Estimating Model Prediction Accuracy: A stochastic Approach to Ecosystem Modeling

Fedra, K. (1980). Estimating Model Prediction Accuracy: A stochastic Approach to Ecosystem Modeling. IIASA Working Paper. IIASA, Laxenburg, Austria: WP-80-168

Preview

Text WP-80-168.pdf Download (928kB) | Preview

Abstract

Ecosystems are, as a rule, characterized by a large behavioral repertoire showing a high degree of structural variability and complex control mechanisms such as adaptation and self-organization. Our quantitative understanding of ecosystems behavior is generally poor, and field data are notoriously scarce, scattered, and noisy. This is most pronounced on a high level of aggregation where considerable sampling errors are involved. Also, no well established and generally accepted ecological theory exists, so that an operational ecosystem model consists of many more arbitrary, simplifying assumptions (more often than not implicitly hidden in process descriptions) than properties measurable in the field. Consequently, predictions of future systems behavior under changed conditions -- a most desirable tool for environmental management -- cannot be precise and unique in a deterministic sense. Rather, it is essential to estimate the levels of model reliability and the effects of various sources of uncertainty on model prediction accuracy. A concept of allowable ranges for model data-input and expected model response, explicitly including uncertainty in the numerical methods, is proposed. Straightforward Monte Carlo simulation techniques are used, and the approach is exemplified on a lake ecosystem eutrophication problem. The method attempts to predict future systems states in terms of probability distributions, and explores the relations of prediction accuracy to data uncertainty and systems variability, the time horizon of the prediction, and finally the degree of extrapolation in state- and input-space relative to the empirical range of systems behavior. The analysis of almost 100,000 model runs also allows some conclusions on model sensitivity, and some desirable model properties in light of prediction accuracy are identified.

Actions (login required)

View Item