Türkiye Tarımsal Araştırmalar Dergisi (Sep 2015)

Predicting of Soil Aggregate Stability Values Using Artificial Neural Networks

  • Tuğrul YAKUPOĞLU,
  • Ayfer Özge ŞİŞMAN,
  • Recep GÜNDOĞAN

Journal volume & issue
Vol. 2, no. 2
pp. 83 – 92

Abstract

Read online

In studies conducted in the field of agriculture forecasting engineering today has come to an important point and forecasting of artificial neural networks (ANN) use has become wide spread. In this study, wet aggregate stability (WAS) depending on the seasonal variation has been investigated whether it can be estimated or not using ANN pilot area soils located in, Avşar Campus of Kahramanmaraş Sütçü İmam University. Selected based on the results of the statistical evaluation of soil properties were used as independent variables and predictive ANN’s have been developed to WAS. In Network training WAS values that are the closest to the actual have tried to reach by using twelve different learning algorithms. Used in the training of these different Back-propagation algorithms’ performances were evaluated using coefficient of determination (R2), the square root of the mean square error (RMSE) and mean absolute percentage error (MAPE). R2 values 0.55-0.99, RMSE values 2.12-11.33 % and MAPE values 3.55-20% of the created ANNs through education different algorithms has changed in the ranges. ANNs was developed when R2 was compared with each other on the basis of criteria networks in terms of estimation power to WAS, R2 values were found above 0.97 of BFGS algorithm with the exception of trained network of all. On the other hand, created ANNs when evaluated according to the criteria of the RMSE has been reached to result that most successful network was developed network with RP's algorithm (12.2%) and the most failed network was developed network with BFGS (11.33%) algorithm. Considering the MAPE indication of the forecasting power, the highest network with OSS algorithm (3.55%) the trained with ANN and forecasting power has been the lowest trained ANN with BFGS algorithm (20%). The results obtained indicate that when ANN was created using the correct training algorithm can be used in the estimation WAS.

Keywords