Modeling Soil Penetration Resistance Using Regression, Artificial Neural Network and Gene Expression Programming

• Published in: Majallah-i āb va khāk Vol. 38; no. 2; pp. 269 - 283
• Main Authors: Sh. Asghari, M. Hasanpour Kashani, H. Shahab Arkhazloo
• Format: Journal Article
• Language: Persian
• Published: Ferdowsi University of Mashhad 01.05.2024
• Subjects: estimation; fandoghloo region; gravitational water content; readily variables; soil resistance
• ISSN: 2008-4757; 2423-396X
• DOI: 10.22067/jsw.2024.86792.1385

IntroductionThe penetration resistance (PR) of the soil shows the mechanical resistance of the soil against the penetration of a conical or flat probe; it is important in terms of seed germination, root growth and tillage operations. In general, if the PR value of a soil exceeds 2.5 MPa, the growth and expansion of roots in the soil will be significantly limited. The direct measurement of PR is also a laborious and costly task due to instrumental errors. Therefore, it is useful the use of different models such as multiple linear regression (MLR), artificial neural network (ANN) and gene expression programming (GEP) to estimate PR through easily accessible and low-cost soil characteristics. The objectives of this research were: (1) to obtain MLR, ANN and GEP models for estimating PR from the easily accessible soil variables in forest, range and cultivated lands of Fandoghloo region of Ardabil province, (2) to compare the accuracy of the aforementioned models in estimating soil PR using the coefficient of determination (R2), root mean square error (RMSE), mean error (ME) and Nash-Sutcliffe coefficient (NS) criteria. Materials and MethodsDisturbed and undisturbed samples (n = 80) were nearly systematically taken from 0-10 cm soil depth with nearly 50 m distance in forest (n = 20), range (n = 23) and cultivated (n = 37) lands of Fandoghloo region of Ardabil province, Iran (lat. 38° 24' 10" to 38° 24' 25" N, long. 48° 32' 45" to 48° 33' 5" E) in summer 2023. The contents of sand, silt, clay, CaCO3, pH, EC, bulk (BD) and particle density (PD), organic carbon (OC), gravimetric field water content (FWC), mean weight diameter (MWD) and geometric mean diameter (GMD) were measured in the laboratory. Relative bulk density (BDrel) was calculated using BD and clay data. Mean geometric diameter (dg) and geometric standard deviation (σg) of soil particles were computed by sand, silt and clay percentages. The penetration resistance (PR) of the soil was measured in situ using cone penetrometer (analog model) at 5 replicates. Data randomly were divided in two series as 60 data for training and 20 data for testing of models. The SPSS 22 software with stepwise method, MATLAB and Gene Xpro Tools 4.0 software were used to derive multiple linear regression (MLR), artificial neural network (ANN) and gene expression programming (GEP) models, respectively. A feed forward three-layer (2, 5 and 6 neurons in hidden layer) perceptron network and the tangent sigmoid transfer function were used for the ANN modeling. A set of optimal parameters were chosen before developing a best GEP model. The number of chromosomes and genes, head size and linking function were selected by the trial and error method, as they are 30, 3, 8, and +, respectively. The rates of genetic operators were chosen according to literature studies. The accuracy of MLR, ANN and GEP models in estimating PR were evaluated by coefficient of determination (R2), root mean square error (RMSE), mean error (ME) and Nash-Sutcliffe coefficient (NS) statistics. Results and Discussion The studied soils had clay loam (n = 11), sandy clay loam (n = 6), sandy loam (n = 12), loam (n = 13), silty clay loam (n = 14), silty clay (n = 1) and silt loam (n = 23) textural classes. The values of sand (13.14 to 64.79 %), silt (21.11 to 74.96 %), clay (2.95 to 42.18 %), OC (1.01 to 7.17 %), FWC (11.58 to 50.47 mass percent), BD (0.84 to 1.43 g cm-3) and PR (1.03 to 5.83 MPa) showed good variations in the soils of the studied region. There were found significant correlations between PR with FWC (r = - 0.45**), silt (r = - 0.36**) and σg (r = 0.36**). Due to the multicollinearity of silt with σg (r = -0.84**), the σg was not used as an input variable to estimate PR. Generally, 3 MLR, ANN and GEP models were constructed to estimate PR from measured readily available soil variables. The results of MLR, ANN and GEP models showed that the most suitable variables to estimate PR were FWC, silt and BDrel. The values of R2, RMSE, ME and NS criteria were obtained equal 0.44, 1.19 MPa, 0.19 MPa and 0.36, and 0.92, 0.41 MPa, -0.05 MPa and 0.92,  0.79, 0.91 MPa, 0.13 MPa, 0.63 for the best MLR, ANN and GEP models, respectively. The former researchers also reported that there is a negative and significant correlation between PR with FWC. Conclusion The results indicated that field water content (FWC), silt and relative bulk density (BDrel) were the most important and readily available soil variables to estimate penetration resistance (PR) in the studied area. According to the lowest values of RMSE and the highest values of NS, the accuracy of ANN models to predict soil PR was higher than MLR and GEP models in this research.