A comparison of neural network and process-based models for vegetation distribution under global climate change

Ecological models based on physical, chemical, and biological processes are diicult to develop and calibrate. Neural network methods provide an attractive alternative where the model is constructed automatically. We compared the MAPSS global vegetation model to a collection of neural network models trained on a data set covering current climate and vegetation distribution in the coterminous United States. Completely automated methods were employed to calibrate MAPSS and to train the neural networks. The models were then compared on two tasks: prediction of current vegetation distribution given current climate and prediction of future vegetation distribution given future 2CO 2 climates produced by the OSU, GFDL, and UKMO global climate models. For predicting under current climate, the neural network models are more accurate than MAPSS. Unfortunately, the predictive accuracy of the models under future climate models cannot be measured without knowing future vegetation distribution. However, an upper bound on the accuracy can be computed using a statistic called the classiication scatter. On future climate scenarios, the neural network models exhibit much higher classiication scatter than MAPSS, which means that the highest accuracy that they could achieve is substantially less than the highest potential accuracy of MAPSS. This suggests|but does not prove|that process-based models such as MAPSS will make better predictions than neural network models for future climate scenarios.