Compared with traditional soil nickel (Ni) content determination methods, visible near-infrared spectroscopy (Vis–NIR) technology can achieve the fast and non-destructive prediction of soil Ni content. However, Vis–NIR spectroscopy data are susceptible to environmental factors during the collection process; thus, it is necessary to perform appropriate preprocessing operations before modeling to improve the data quality and modeling accuracy. In this study, we focus on the polluted farmland around the gold mine in Mojiang Hani Autonomous County, Yunnan Province. First, Savitzky–Golay smoothing was applied to the spectrum (R), and then the impact of using second-order derivative processing (R’’) on modeling accuracy was investigated. The potentials of recurrent neural networks (RNNs), random forests (RFs), and partial least squares regression (PLSR) to predict soil Ni content were explored. The results indicated the following: (1) The model established by transforming R with second-order derivatives has shown a clear improvement in prediction accuracy. The use of second-order derivatives helps eliminate the effect of baseline drift on the spectra and also serves to remove noise and amplify differences between features. (2) RNN has the best performance among the three modeling methods, followed by RF and PLSR. Owing to the complex nonlinear relationships between spectral data, RNN has a greater advantage in coping with this situation, and RF has a limited capability to deal with this situation, which PLSR as a linear model does not have. (3) The best model for predicting soil Ni content in this study is R’’-RNN, which has high prediction accuracy and generalization ability. Its validation set root mean square error (RMSE), coefficient of determination (R2), relative analysis error (RPD), and ratio of performance to interquartile range (RPIQ) are 116.81 mg/kg, 0.85, 2.55, and 4.05, respectively. This study provides a new reference approach for monitoring heavy metals in contaminated farmland soil around gold mines.