Abstract: Estimating the total nitrogen (TN) content of soil accurately and rapidly is the guarantee to promote formula fertilization development. This research selected 30 point locations randomly from different regions. Then the topsoil layer (0-30 cm), subsoil layer (＞30-48 cm) and ground layer (＞48-60 cm) of each point were chosen to get soil samples. And these samples were used for all the subsequent experiments. The near infrared spectral absorbance of soil samples with different nitrogen contents was measured using the Fourier spectrum analyzer MATRIX-I. At the same time, the TN content of each sample was measured using Kjeldahl method in the laboratory. Then the absorbance spectral characteristics of soil samples from different layers were analyzed including the change laws of soil moisture and TN content from layer to layer. The first order differential processing was conducted among the 90 soil samples' original spectral absorbance. Then the correlation analyses were done between the TN content and the original or differential spectral data respectively. From the results of correlation coefficient between differential spectra and TN content, 1387, 1496, 1738, 1875, 2116 and 2314 nm were selected as sensitive wavebands finally. The sensitive wavebands were used to establish the multiple linear regression (MLR) model, the model based on back propagation (BP) neural network and the BP neural network prediction model optimized by the genetic algorithm to predict the soil TN content. The results were showed as below. For MLR model the accuracy of calibration process was high, while in predicting process, the modeling accuracy decreased with the increase of soil depth. For the model based on the BP neural network, it had good universality in predicting the TN content in the different layers of soil. To some extent, this method improved the prediction ability under the background of high moisture, while the model accuracy was yet lower. For the BP neural network model optimized by the genetic algorithm, the R2 of the calibration process reached 0.883, and the root mean square error (RMSE) of calibration was 0.0278 mg/kg. The R2 of prediction for topsoil layer reached 0.716, and the RMSE of prediction was 0.031 mg/kg. The R2 of prediction for subsoil layer was 0.801, and the RMSE of prediction was 0.030 mg/kg. The R2 of prediction for ground layer reached 0.667, and the RMSE of prediction was 0.033 mg/kg. Compared with the MLR model and the BP neural network model without optimization, the BP neural network model optimized by the genetic algorithm showed a significant improvement in both calibration and predicting accuracy for each soil layer. Therefore, the BP neural network prediction model optimized by the genetic algorithm has obvious advantages for soil TN content prediction.