Assessing Performance of Multivariate Linear Regression (MLR), artificial neural network (ANN) and Gene Expression Programming (GEP) in estimating soil

Assessing Performance of Multivariate Linear Regression (MLR), artificial neural network (ANN) and Gene Expression Programming (GEP) in estimating soil properties

Abstract
Background and Objectives: With the emergence of computers and geographic information system (GIS), as well as access to spatial digital data, different methods of data mining, modeling and estimation of soil properties found their place in soil sciences. Data mining of soil properties using computer-based statistical methods uncovers hidden patterns in the database which ultimately leads to the fitness of a model for estimation of soil properties. These methods can be used in the scorpan equation. Two main components of scorpan model include environmental variables and learning program. In the present study, three different methods including multiple linear regression (MLR), artificial neural network (ANN) and Gene Expression Programming (GEP) as “f’ function in scorpan model were evaluated and compared in estimating of soil properties using auxiliary data such as vegetation data, topography and remote sensing data.
Material and methods: The study area with an area of 1225 ha was located in Bajgiran rangelands, Khorasan Razavi province, Iran. In order to investigate vegetation cover and soil 137 units were investigated in which 3-5 plots were selected with a distance of 10 meters apart along an accidental transect, and plant species names and numbers besides vegetation percentage were recorded. Next, one soil sample was taken from each transect (Totally 137 soil sample). Train attributes derived from digital elevation model; different bands derived from the ETM and used for computing spectral indices; and plant diversity indices were calculated using Simpson and Shannon-Wiener. These obtained parameters were used as covariate in estimating calcium carbonate equivalent, clay, density, nitrogen, carbon, sand, silt and saturated moisture capacity. Data deduction was done by PCA analysis to deduct the number of input data for ANN and GEP models, and finally, Normalization and standardization were carried out on the data.
Results: The results obtained from the evaluation of three numerical methods based on root mean square error (RMSE), mean bias error (MBE) and coefficient of determination (R2) showed that ANN model had the highest accuracy in estimating soil properties, given the higher coefficients of determination for calcium carbonate equivalent , clay, density, nitrogen, carbon, sand, silt and saturated moisture capacity with the values of 0.72, 0.46, 0.69, 0.67, 0.77, 0.62, 0.7 and 0.85, respectively, moreover, lower RMSE with the values of 7.46, 4.46, 0.08, 0.03, 0.27, 5.6, 3.5 and 3.4, respectively. ANN could explain 60-85 percent of variability of soil properties, among which the best estimates were for saturated moisture capacity and soil organic carbon with R2 = 0.85 and R2 = 0.77, respectively.
Conclusion: Evaluating the estimation of soil properties through three numerical models introduced ANN as the most accurate model. ANN validation results showed that mean bias error (MBE) for estimated soil properties were close to zero, and this confirms that the fitting has been created unbiased by model. Furthermore, the low RMSE of model verified accurate estimation of soil variables. The results also indicate that GEP had higher accuracy than the linear regression method for most soil properties.