Estimation models of soil organic matter (SOM) and soil total nitrogen (TN) were established based on NIR spectroscopy and BP neural network A total of 150 soil samples were collected from the tested farm, and the NIR spectra of all soil samples were measured. First, data pretreatment was performed for each sample with the method of locally weighted scatter plot smooth filtering. Then the box plot analysis for the measured SOM data and TN data were conducted separately and the information about the shape, location, and distribution of the target data was obtained. The variance between the SOM data was very small, and most of them were concentrated on the median. This was also observed from TN data. Thus clustering analysis was carried out for the target parameters of the soil samples so that the original dataset with 150 spectra was clustered to 50 groups. For each group, the average of spectral data was calculated at every wavelength to obtain a new spectrum. The new spectrum was calculated with natural logarithm and normalized, which was taken as a new sample. Principal component analysis (PCA) was executed for 50 new samples and the principal components with over 99.98% of cumulative proportion of correlation matrix were extracted to establish BP neural network. According to the analysis result of SOM content, the calibration accuracy of the model was 0.999, and the validation accuracy reached to 0.854. According to the analysis result of the soil TN content, the calibration accuracy of the model was close to 1, and the validation accuracy reached 0.808. The result shows that the smooth filter can weaken the noise in the data, expose the data features, provide a reasonable starting approach for parametric fitting; and improve the prediction accuracy; It is feasible and practical to estimate soil parameters by using BP neural network with the prediction accuracy of 0.854 (SOM) and 0.808 (TN); Compared to the other prediction modeling method, the BP neural network model has higher robustness and better fault tolerance, and the model accuracy would not be affected by the several outline samples when the number of samples is large enough.