PREDICTION OF CROP YIELDS ACROSS FOUR CLIMATE ZONES IN GERMANY: AN ARTIFICIAL NEURAL NETWORK APPROACH

This paper shows the ability of artificial neural network technology to be used for the approximation and prediction of crop yields at rural district and federal state scales in different climate zones based on reported daily weather data. The method may later be used to construct regional time series of agricultural output under climate change, based on the highly resolved output of the global circulation models and regional models. Three 30-year combined historical data sets of rural district yields (oats, spring barley and silage maize), daily temperatures (mean, maximum, dewpoint) and precipitation were constructed. They were used with artificial neural network technology to investigate, simulate and predict historical time series of crop yields in four climate zones of Germany. Final neural networks, trained with data sets of three climate zones and tested against an independent northern zone, have high predictive power (0.83 < R² < 0.9). Hindcasts, based on a 25-year training period and independent weather data of a 5 (3)-year future have a relative root mean square error of less than 9%. The model approximates and predicts historical reported yields in an area with a wide range of climatic variance and heterogeneous soil conditions. Mean temperatures during growing seasons ranged from 8.7° C (10.4°) to 19.3° C (21.1°) for April - July (May – September) and precipitation from 73 mm (141) to 548 mm (1016). The output of general circulation models and dynamical crop growth models can easily be integrated to simulate impacts of climate change.