Prediction of functional characteristics of ecosystems: A comparison of artificial neural networks and regression models

We tested the potential of artificial neural networks (ANNs) as predictive tools in ecology. We compared the performance of ANNs and regression models (RM) in predicting ecosystems attributes, with special emphasis on temporal (interannual) predictions of functional attributes of the ecosystem at regional scales. We tested the predictive power of ANNs and RMs using simulated data for six functional traits derived from the seasonal course of the normalized difference vegetation index. (NDVI): the annual integral of the NDVI curve (NDVI-I), the maximum (MAX) and minimum (MIN) NDVI, the date of the MAX NDVI (DM) and the date of start (SGS) and end (EGS) of the growing season. For one of these: traits (NDVI-I), we also generated. set of data that incorporated the effects of the state of the system in previous years (inertial effects). Even simple non-linearities in the actual functional form of the relationship between environmental variables and ecosystem attributes preclude a precise prediction of these attributes when the rules are nor explicit. That was evident for predictions based on both ANNs and RMs under absolutely deterministic conditions (error-free scenario). Non-linearities in the simulated traits of the NDVI curve derive from multiplicative terms in the models. Under the presence of these non-linear terms, a different aggregation of the driving variables (monthly vs. annual or quarterly climatic data) reduce substantially the ability of both RMs and ANNs to predict the independent variable. For the six traits analyzed, the ANNs were able to make better predictions than RMs. The correlation between observed and predicted values of each of the six traits considered was higher for the ANNs than for the RMs. ANNs showed clear advantages to capture inertial effects. The ANN used was able to use previous year information on climate to estimate current year NDVI-I much better than the RM that used the same input information. (C) 1997 Elsevier Science B.V.