Soil surface shear strength is an important parameter for prediction of soil erosion, but its direct measurement is difficult, time-consuming and costly in the watershed scale. This study was done to predict soil surface shear strength using artificial neural networks (ANNs) and multiple linear regression (MLR) and to rank the most important soil and environmental attributes affecting the shear strength. A direct shear box was designed and constructed to measure in situ soil surface shear strength. The device can determine two soil shear strength parameters i.e. cohesion (c) and angle of internal friction (φ). The study area (3500 km2) was located in Semirom region, Isfahan province. Soil surface shear strength was determined using the shear box at 100 locations. Soil samples were also collected from 0-5 cm layer of the same 100 locations at which the surface shear strength was measured using the shear box. Particle size distribution, fine clay content, organic matter content (OM), carbonate content, bulk density and gravel content were determined on the collected soil samples. Normalized difference vegetation index (NDVI), the type of land use and geology were also determined. The MLR and ANNs were used to model/predict soil surface shear strength (c and φ). In order to compare the modeling methods, coefficient of determination and root mean square errors were used as efficacy indices. The results showed that ANN models were more feasible in predicting soil shear strength parameters than MLR models due to capability of ANN models in deriving nonlinear and complex relations between the parameters. Results of sensitivity analysis for ANN models indicated that NDVI, bulk density and fine clay content are more effective parameters in predicting c in the studied region. Also it was found that sand content, bulk density and NDVI were more effective parameters and OM/clay ratio and organic matter content were less effective parameters in predicting φ in the region.