Monitoring total nitrogen content rapidly and accurately in farmland soils can significantly improve the efficiency of soil fertility diagnosis and evaluation efficiency. Traditional methods for measuring total soil nitrogen have disadvantages such as time-consuming, high cost, and environmental pollution, while the quantitative method of total soil nitrogen based on spectroscopic principles overcomes the disadvantages of traditional measurements. Mid-infrared (MIR) spectroscopy has morebands and information than visible-near infrared (VNIR) spectroscopy, and how to use MIR spectroscopy to monitor soil total nitrogen content has not been systematically studied in China. In order to explore the feasibility of mid-infrared spectroscopy for soil total nitrogen monitoring,we selected 246 farmland soil samples in the southern Xinjiang region as the research objects and used total nitrogen content and mid-infrared spectral reflectance data measured in the laboratory to analyze the differences mid-infrared spectral characteristics of soil samples with different total nitrogen content. Firstly, the dimension of spectral data was reduced by the principal component analysis (PCA) and successive projections algorithm (SPA) and then used four modeling methods including the partial least squares regression (PLSR), support vector machine (SVM), random forest (RF) and back propagation neural network (BPNN) to construct the quantitative inversion model of soil total nitrogen content based on the full-band and dimension-reduced data, respectively. The results showed that: (1) the spectral reflectance of soil in the mid-infrared band increased with the increase of total nitrogen content with a significant absorption valley near 3 620, 2 520, 1 620 and 1 420 cm^-1. The correlation between soil spectral reflectance and total nitrogen content could improve significantly after the maximum normalization of mid-infrared spectral data. (2) Comparing the two data dimension reduction methods, PCA and SPA reduced the number of model variables by 99.8% and 97.5% respectively. However, the prediction accuracy of the model established with the eight principal components extracted by PCA as independent variables were generally higher than that of the corresponding model of SPA. Therefore, the modeling with the principal components extracted by PCA was more suitable for constructing the soil total nitrogen model. (3) In the modeling set, the PLSR and SVM models had the highest accuracy in full-band modeling, but the modeling efficiency was low due to alarge number of modeling variables. However, based on the RF and BPNN models, using the data after dimension reduction by PCA and SPA for modeling respectively, while maintaining the comparative accuracy, the modeling efficiency can be significantly improved. In the prediction set, the BPNN model based on PCA dimension-reduced data had the highest prediction ability, with R² and RMSE of 0.78 and 0.12 g·kg^-1, RPD and RPIQ of 2.33 and 3.54, respectively, indicating that the model had great prediction ability. The study results can provide some reference values for the rapid estimation of total nitrogen content in farmland soil.