Nitrite is a common water quality pollutant andis the main source of wastewater, fertilizer, and sewage treatment plants. The size of nitrite concentration in water quality is an important indicator to assess the health of water bodies. Still, the traditional method of nitrite concentration detection is complicated to operate. It easily interferes with the detection environment, which can not intuitively and accurately reflect the health of water quality. To explore a new way to assess the health of water bodies, this paper uses the IPSO-BPNN model to predict the concentration of nitrite transmission spectral data. Ten concentrations of nitrite standard solutions (0.02, 0.04, 0.06, 0.08, 0.10, 0.12, 0.14, 0.16, 0.18, and 0.20 mg·L^-1) are first selected, and the ten concentrations of nitrite solutions are scanned at the same time intervals by using the OCEAN-HDX-XR micro spectrometer, The spectral transmittance of the spectral data is obtained by white board calibration to obtain spectral transmittance values for the spectral data. Two preprocessing methods, maximum-minimum normalization, and mean-centering, are used to unify the spectral data into uniform dimensions and centroids, making the spectral data comparable and interpretable among different samples. Due to the high dimensionality of the original spectral data, kernel principal component analysis is used for data dimensionality reduction, and six principal components representing 97.94% of the original data information are selected for the training of the IPSO-BPNN model. When predicting nitrite concentration, the original particle swarm optimization algorithm is improved by introducing adaptive learning factor and inertia weight updating formulae and particle population diversity guiding strategy, and learning rate adaptive formulae are introduced based on the BP neural network to improve the algorithm’s performance. By comparing the change curves of function fitness values for iterations performed under different particles, 30 iterations using 100 particles are chosen to find the optimal weight and bias combinations. The results show that the coefficient of determination of the IPSO-BPNN prediction model is 0.983 760, the root-mean-square error is 0.007 320, and the average absolute error is 0.004 705, which is a better fit compared with the current Random Forest, Linear Regression, BP-ANN, PSO-BPNN, and PSO-SVR models that have better prediction performance and higher accuracy. Based on these results, a hyperspectral water quality nitrite concentration prediction method based on the IPSO-BPNN model is proposed, providing a new idea for assessing water body health.