Using the ANN different algorithms in modeling soil temperature on different depths in terms of some climatic information

Abstract

In this research, the neural network model was applied to estimate the soil temperatures on various depths and with different times periods in terms of Air_temperature, sunshine and radiation for a given day of the year using the BackPropagation Network (BPN). Different algorithms were used including; Gradient Descent algorithm (GD) , Gradient Descent with Momentum algorithm (GDM) , Conjugate Gradient algorithm (CFG) , Quasi-Newton algorithm (BFGS) and Levenberg-Marquardt algorithm (LM) . Data was obtained from the soil and water research department / Nineveh province for the period from 1980 –1983 was used, and date included daily soil temperature on depths of (5, 10, 20, 30, 50 and 100) CM and for three time periods at (9, 12, and 15 hours) for cultivated and uncultivated soils. The data of two years was used to develop the models and the data of 1983 was used to evaluate the models and comparing the outputs with the data measured. Also, the R2 scales and the root of mean square error (RMSE) were used in determining the match of the measured data and the neural network outputs to select the best predictive model out of the applied models. Results showed that the (LM) algorithm proves to be highly efficiency in improving a good predictive model for temperature of various depths of soil. In addition, this algorithm is also considered the best and fastest algorithm for constructing such models for any daily cultivated and uncultivated soil at different depths if Air temperature , sunshine and radiation are available at the three measuring periods for any day of the year .