Prediction of Daily Soil Temperatures with Artificial Neural Network

Soil temperature is one of the key parameters affecting most hydrologic and agricultural processes. Therefore, its measurement and prediction is very crucial. Since soil temperature is measured only in the synoptic meteorological stations, lack or shortage of data is the major challenge in many agricultural studies. In this study, soil temperature data were predicted at three different depths of 5, 10 and 30 cm by an Artificial Neural Networks using meteorological parameters recorded at Shiraz synoptic stations during the period 2000-2008. Due to the large number of variables used in this study to predict soil temperature, identifying the more effective variables could improve the results. Therefore, using a multivariate statistical technique of principal component analysis (PCA), which effectively reduces the number of variables and inputs the effective variables to the network, the soil temperature was predicted by (PCA-ANN).  At first, PCA was used to reduce the input variables and 8 meteorological variables were altered to 8 main components. The first four principal components accounted for over 99% of the total variance. In order to evaluate ANN and PCA-ANN models, correlation coefficient (r), Root Mean Square Error (RMSE) and Mean Bias Error (MBE) were used. Results showed that the statistical parameter values (verification period) with 0.98, 1.61 and 0.2 for r, RMSE, MBE, respectively, had the best results on soil temperature at 5 cm depth for PCA-ANN model. The results indicated the importance of preprocessing the variables by PCA. The PCA-ANN model in comparison with the ANN model had an easier structure, the ability of faster training, and more accurate results.