Freeze prediction for specific locations using artificial neural networks

Artificial neural networks (ANNs) were developed to predict air temperature in I h increments from I to 12 h in the future. Weather data for model development and evaluation for three locations in Georgia (Fort Valley, Blairsville, and Alma) were obtained from the Georgia Automated Environmental Monitoring Network (AEMN). The data consisted of observations of meteorological variables such as air temperature, relative humidity, wind speed, rainfall, and solar radiation. The critical inputs for each model were determined by developing ANNs that used them in various input combinations and observing their effect on the accuracy of the ANN predictions. The results showed that of the meteorological variables considered, only rainfall was not useful in generating air temperature predictions. The optimal duration of prior data ranged from 2 h to 6 h, depending on the period of prediction. The mean absolute error (MAE) increased as the period of prediction got longer The MAE of the evaluation dataset for predicting temperature 1 h in advance was 0.6 degrees C for Fort Valley, 0.7 degrees C for Blairsville, and 0.6 degrees C for Alma. The corresponding MAE values for a 12 h prediction were 2.4 degrees C, 3.0 degrees C, and 2.6 degrees C. Further efforts will be directed to developing general ANNs based on data from multiple locations. The availability of decision support systems that incorporate localized temperature predictions for use by fruit growers could have a positive impact on frost damage protection.