
Prediction of soil macronutrients using fractal parameters and artificial intelligence methods
Author(s) -
Ali Asghar Zolfaghari,
Meysam Abolkheiryan,
Ali A. Soltani-Toularoud,
Ruhollah TaghizadehMehrjardi,
Amanuel Oqbit Weldeyohannes
Publication year - 2020
Publication title -
spanish journal of agricultural research
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.337
H-Index - 36
eISSN - 2171-9292
pISSN - 1695-971X
DOI - 10.5424/sjar/2020182-15460
Subject(s) - silt , soil science , soil test , artificial neural network , support vector machine , fractal dimension , soil texture , lime , mathematics , soil ph , environmental science , fractal , soil water , computer science , machine learning , geology , paleontology , mathematical analysis
Aim of study: To evaluate artificial neural networks (ANN), and k-Nearest Neighbor (k-NN) to support vector regression (SVR) models for estimation of available soil nitrogen (N), phosphorous (P) and available potassium (K).Area of study: Two separate agricultural sites in Semnan and Gorgan, in Semnan and Golestan provinces of Iran, respectively.Material and methods: Complete data set of soil properties was used to evaluate the models’ performance using a k-fold test data set scanning procedures. Soil property measures including clay, sand and silt content, soil organic carbon (SOC), electrical conductivity (EC), lime content as well as fractal dimension (D) were used for the prediction of soil macronutrients. A Gamma test was utilized for defining the optimum combination of the input variables.Main results: The sensitivity analysis showed that OC, EC, and clay were the most significant variables in the prediction of soil macronutrients. The SVR model was more accurate compared to the ANN and k-NN models. N values were estimated more accurately than K and P nutrients, in all the applied models.Research highlights: The accuracy of models among the test stages illustrated that using a single data set for investigation of model performance could be misleading. Therefore, the complete data set would be necessary for suitable evaluation of the model.