Prediction of Some Difficult-to-measure Soil Characteristics Using Regression Pedotransfer Functions and Artificial Neural Network in Kerman Province

Measurement of some important soil characteristics may be difficult, time-consuming, and expensive. Thus, it is helpful to predict these properties using easily-available soil properties. These relationships and/or functions are called pedotransfer functions (PTFs). This study was conducted to derive PTFs for estimating field capacity (FC), permanent wilting point (PWP), and cation exchange capacity (CEC) of soils in Kerman Province. Hundred soil samples (0‒30 cm layer) were collected from different locations in Kerman Province including: Kerman, Bardsir, Rafsanjan, Shahre-Babak, Sirjan and Orzoueiyeh of Baft. Then, FC, PWP, CEC, clay, silt, sand, carbonate, organic matter and gypsum contents of the soils were measured. In the regression method, clay, sand, and gypsum contents significantly affected the FC prediction, whereas clay content entered as effective input in the derived model for PWP, and clay and organic matter contents had significant effects on the CEC. Coefficients of determination (i.e. R2) of 0.86, 0.45 and 0.94 were calculated for FC, PWP, and CEC regression models, respectively. The best PTFs were obtained by artificial neural network (ANN) for FC, PWP and CEC with 6 hide layers and including all the input variables (R2 values of 0.98, 0.93 and 0.99, respectively). The accuracy of ANN predictions was greater than that of regression method. Results revealed that regression models can be applied with acceptable accuracy if a few easily-available characteristics are measured. The ANN method presented highly accurate results when the number of known easily-available characteristics increased. The accuracy of ANN decreased with reducing the number of inputs.