The development of accurate calibration models is a crucial prerequisite for the successful application of near-infrared (NIR) spectroscopy for the determination of N contents in the reclaimed mine soils. The objective of this study was to compare the performance of calibration models developed with the artificial neural network (ANN) and partial least squares (PLS) methods. The topsoil (0 – 20 cm) samples (n = 92) were taken at the reclaimed spoil heap of Bełchatów lignite mine (Poland, 51°22′07″N, 19°21′24″E) and measured for the N content with Kjeldahl method. The samples were finely ground and their NIR (1000 nm – 2500 nm) were recorded. The calibration models were developed with 62 samples and then validated with the remaining 30 samples. At the calibration stage the models developed with both methods were very effective with the coefficient of determination (reference vs NIR predicted values) R2>0.90 and the ratio of prediction to deviation (RPD) >4. However, tests of the models with the validation sample set yielded somewhat worse results with R2 = 0.79 and RPD = 2.32 for PLS model and R2 = 0.80 and RPD = 2.19 for ANN model. The PLS model was slightly more accurate than ANN model as indicated by lower standard error of prediction values (0.16 mg g-1 and 0.17 mg g-1 for PLS and ANN models, respectively). However, the regression coefficient of linear regression (reference vs NIR predicted values) was better for the ANN model (a = 0.84) than for the PLS model (a = 0.73). The coefficients of determination (reference vs NIR predicted values) were very similar in both methods (R2 = 0.79 and 0.80, for PLS and ANN respectively). The results indicated that both tested calibration models – ANN and PLS gave comparable results in terms of prediction accuracy. Hence, both methods can be used to create predictive models of the N content in the mine soils.