Comparison of pedotransfer functions based on machine learning methods to estimate soil moisture at field capacity and permanent wilting point (Case study: Ravansar District, Kermanshah Province)

The physical properties of the soil, which cannot be easily measured, play an important role in the design of irrigation and drainage systems. Since the direct measurement of these characteristics is time-consuming and expensive, therefore, to estimate these parameters, most researchers use indirect methods such as transfer functions. This research aims to investigate and determine the best model for estimating soil moisture content in Field capacity (FC) and Permanent Wilting Point (PWP) using easily measured soil characteristics and pedotransfer functions in the R software environment and choosing the most suitable model for the soils of the Ravansar region in Kermanshah province.

In this research, the easily measurable properties of soil were used as input variables for five transfer functions of the multivariable linear, artificial neural network, Cubist, random forest, and support vector machine. At first, in the study area, the location of 120 profiles was determined using the Latin hypercube method. In these observation points, soil profile was dug and studied and samples were taken from its horizons. Then, laboratory analysis including measurement of electrical conductivity, pH, calcium carbonate equivalent, organic carbon, and percentage of sand, silt, and clay was performed on soil samples. Based on the range of changes of these characteristics, especially the soil texture, 75 surface soil samples and 33 soil samples from ten different soil profiles were selected. PWP measurement was performed on 33 samples and FC measurement was performed on surface and depth samples, i.e., 108, and then modeling operations were performed on them. Root mean square error (RMSE), mean absolute error (MAE), and R2 indices were used to evaluate the models.

The results showed that the accuracy of pedotransfer functions in estimating PWP is higher than FC (R2 and RMSE values of the Cubist model for PWP are 0.81 and 0.054 and for FC are 0.53 and 0.085, respectively). Also, the results for FC showed that between the models, the Cubist and the artificial neural network have low MAE (0.066 and 0.068) and RMSE (0.085) and high R2 (0.53 and 0.54) respectively, compared to other models.

The overall results showed that Cubist, Artificial Neural Network, and Random Forest models with lower error and higher R2 have higher efficiency for soil moisture estimation in FC than other models. The results showed that the Cubist and random forest models were the best models for estimating moisture at the PWP in terms of comparing the coefficient of determination. This research showed the importance of using new machine learning methods in studies related to soil transfer functions to estimate difficult-to-measure soil properties. Also, the results of this research are acceptable for a wide range of plains in Kermanshah province, which has similar soil formation conditions to the Ravansar region.