[Objective] The optimal model for predicting total phosphorus concentration in water bodies was studied in order to provide support for accurate, real-time, and efficient monitoring of water resources. [Methods] Water quality samples were collected in 2021 from the Wuhan-Anhui region in the middle and lower reaches of the Yangtze River. Firstly, the collected spectral data of the Yangtze River was preprocessed by both maximum-minimum normalization and mean centering to unify the range and mean point of the data. Kernel principal component analysis (KPCA) was then used to perform dimensionality reduction on the preprocessed spectral data. Six feature vectors were selected based on a variance explanation rate of 99.6% for training the subsequent prediction model. Next, an improved particle swarm optimization (IPSO) algorithm was proposed by introducing an adaptive inertia weight updating formula and a genetic-simulated annealing mutation concept to enhance the optimization ability of the algorithm. The improved particle swarm optimization algorithm was used to optimize the hyperparameter combinations in the support vector regression (SVR) model. The support vector regression model was trained using the output results to predict the total phosphorus concentration. Finally, the test set data were used to predict total phosphorus concentration. [Results] A prediction model for total phosphorus content in water using an improved particle swarm optimization (IPSO) combined with support vector regression (SVR) and spectral dimensionality reduction was proposed. The experimental model achieved an R² of 0.973920, a root mean square error of 0.003012, and a mean absolute error of 0.002105 when predicting total phosphorus concentrations in the Yangtze River water. [Conclusion] The proposed method of using spectral data combined with dimensionality reduction techniques, particle swarm optimization algorithms, and machine learning models was determined to be feasible and effective in total phosphorus concentration measurement. The accuracy and fitting effects of the model were better than the accuracy obtained with several well-performing machine learning models.