Identifying the determinant characteristics influencing soil compactibility indices using neural networks and path analysis Running title: Investigation of Soil Compaction by ANN and PA

Soil compactibility could be quantified via different indices, such as maximum dry bulk density (BDmax) and critical water content (θcritical) in a compaction test. The objective of this study was to determine the soil properties influencing soil compactibility by evaluating pedotranfer functions (PTFs) with respect to their accuracy and function for the prediction of BDmax and θcritical using linear regression and artificial neural network (ANN) methods. To this end, 100 soil samples were collected from the topsoil (0−30 cm) of arable and virgin lands in southeast of Iran. Primary particle size distribution, gypsum, Calcium Carbonate Equivalent (CCE), organic matter (OM) contents, and natural bulk density were used as predictors. Two PTFs were developed using linear multiple stepwise regression: a PTF that estimates BDmax with clay and sand contents and natural bulk density as predictors (R2 = 0.45), and another PTF for the estimation of θcritical employing clay and gypsum contents as predictors (R2 = 0.51). Furthermore, an attempt was made to construct PTFs for the prediction of the BDmax and θcritical utilizing ANNs. High prediction efficiencies were achieved through the ANN models. Generally, when all of the easily available soil properties were included as predictors, the ANN models for the θcritical and BDmax as compared with the results of linear regression method obtained much more accurate estimations. Sensitivity analysis performed in this study via Hill method (shirani, 2017) showed that the most important variable in BDmax prediction-using ANNs is the natural bulk density (BDnatural), followed by sand and clay, CCE, and gypsum contents. The highest sensitivity to clay content belonged to the θcritical and the lowest sensitivity to OM content was observed in the studied soils.