In order to detect soil total nitrogen (TN) content, a portable soil TN detector was developed. The stability, accuracy and precision of the detector were tested with the sixty soil samples collected from Shangzhuang Experiment Station of China Agricultural University. The models were evaluated by the correlation coefficient of calibration (RC), the correlation coefficient of validation (RV), root mean square error of calibration (RMSEC), root mean square error of prediction (RMSEP) and residual prediction deviation (RPD). The results showed that the repeatability error of 940nm, 1050nm, 1100nm, 1200nm, 1300nm, 1450nm and 1550nm were 1.57%, 1.80%, 1.59%, 0.94%, 0.61%, 0.64% and 3.13%, respectively. The correlation coefficients between the absorbance of soil TN detector and that of MATRIX_I Fourier spectral analyzer were 0.971, 0.948, 0.928, 0.873, 0.920, 0.901 and 0.913, respectively. The RC, RV, RMSEC, RMSEP and RPD of the model established by BP neural network was 0.81, 0.80, 0.029, 0.019 and 3.44, respectively. The experiment indicated that the portable soil TN detector performed well. 