In the present study, visible-near infrared reflectance spectroscopy (VNIR) measured in laboratory was evaluated for prediction of the content of As in soils. Calibrations between As and reflectance were developed using cross-validation under partial least squares regression (PLSR). Prediction accuracy was tested via separate validation samples. The reflectance was pre-processed by several techniques like smoothing, multiplicative scatter correction (MSC), Log(1/R), first/second derivative (F/ SD) and continuum removal (CR). The accuracy of prediction was evaluated with three statistics: coefficients of determination (R2), ratio of performance to deviation (RPD), and root mean square error of prediction (RMSEP). The results of calibration, cross-validation and prediction of different pre-processing techniques, spectral resolution and OM content were compared. MSC provided better prediction (Prediction R2 = 0.711, RPD = 1.827, RMSEP = 1.613) than other methods because it removed the effects of light scattering and sample thickness. All the results of different resolution are acceptable (Prediction 0.678 < R2 < 0.711, 1.750 < RPD < 1.827, 1.613 < RMSEP < 1.685). The prediction accuracy of subsets with lower OM content(Prediction R2 = 0.694, RPD = 1.697, RMSEP = 1.644) was better than that with higher content. The study indicates that it is feasible to predict As element in soils using reflectance spectroscopy and the prediction accuracy can be improved by pre-processing. Thus this new rapid and cost-effective technique can be used in the monitoring of soil contamination.