NO₂-N is an important parameter in water bodies and can quickly detect organic pollution parameters. It is of great significance to the assessment of water quality. However, traditional methods are complicated in operation, subject to many interference factors, long measurement time, cannot reflect water quality changes in time, and cannot provide timely and effective early warning. For sudden water pollution incidents, because of the shortcomings of traditional methods, it is of great significance to explore accurate, real-time, and environmentally friendly detection methods for the NO₂-N content in environmental water bodies and drinking water. This experiment is to study the use of superior grade pure reagents to prepare 10 concentrations of NO₂-N nitrogen standard solutions (0.02, 0.04, 0.06, 0.08, 0.1, 0.12, 0.14, 0.16, 0.18 and 0.2 mg·L^-1), using the OCEAN-HDX-XR micro-fiber spectrometer to scan 10 times the transmission spectrum of the NO₂-N solution of each concentration in the range of 181.1~1 023.1 nm. Take the average value as the original transmission spectrum of the NO₂-N solution of each concentration, and then take the NO₂-N content of the solution as the dependent variable and the original transmission spectrum as the independent variable. Use the method of variable feature importance in random forest regression to screen the feature variables. Based on the cross-validation method, the number of the most stable model variables is selected, and the NO₂-N optimization random forest inversion model is established. The results of the study are as follows: (1) The variable explained rate (Var Explained) of the random forest model established by the whole band (Var Explained)=76.49%, and the mean squared residuals (Mean of squared residuals)=0.000 688; In the sensitive band of salt inversion, 195.1 nm has the highest importance value, and the leave-one-out crossover method is used to find that the random forest model has the lowest root mean square error when 19 spectral characteristic variables are used to screen the optimized random forest established by spectral characteristic variables Variable Explained rate (Var Explained)=83.45%, Mean of squared residuals (Mean of squared residuals)=0.000 552. Variable screening effectively reduces the amount of spectral data and provides a basis for the establishment of the optimization model; (3) Model verification of the established model, including the full-band random forest model test set R²=0.820 3, RMSE=0.03, test set R²=0.979 3, RMSE=0.01, optimized random forest model test set R²=0.873 4, RMSE=0.022, test set R²=0.979 8, RMSE=0.008, after comparing the full-band random forest model with the optimized random forest model, it is found that the optimized random forest model test set and test The interpretation and accuracy of the set model are higher than the full-band random forest model, indicating that the optimization method can not only effectively reduce the spectral dimension, but also has positive significance for finding the sensitive band of NO₂-N spectrum and establishing a high-precision NO₂-N inversion model. . Based on the above test results, an inversion method for optimizing the hyperspectral water quality NO₂-N parameters of the random forest model is proposed, which provides a new method for the dynamic detection of water quality NO₂-N parameters.