River wetlands are a critically important type of wetland, with soil playing a vital role in maintaining the stability of riverine wetland ecosystems. Soil carbon, nitrogen, and phosphorus are key nutrient elements supporting soil quality and vegetation growth in wetlands. Utilizing hyperspectral remote sensing data to estimate them is of significant importance for rapid and accurate detection of soil nutrient information in wetlands. Soil particle size, as one of the most important soil properties, has a significant impact on the spectral reflectance of soil samples and is a crucial factor affecting soil structure, cation exchange capacity, and plant nutrient availability. Taking the Shaanxi Yellow River Wetland Provincial Nature Reserve as the study area, 477 surface soil samples were collected from August to September 2022, and after sieving in the laboratory, four different particle-sized soil samples of 1.0 mm, 0.3 mm, 0.2 mm, and 0.1 mm were obtained. Three prediction models, namely partial least squares regression (PLSR), random forest (RF), and Gaussian process regression (GPR), were established for soil carbon, nitrogen, and phosphorus content based on original spectral data and first-order differential transformed spectral data of different particle sizes. The models were compared in terms of modeling R² and RMSR to select the optimal model, and sensitive bands were selected to construct the model for evaluation. The research results showed that: (1) The numerical values of spectral reflectance increased with decreasing soil particle size, and the prediction model for 0.1 mm particle size consistently exhibited better accuracy than other particle sizes; (2) Models for estimating soil organic carbon, total nitrogen, and total phosphorus content based on first-order differential spectra had higher accuracy; (3) Partial least squares regression (PLSR) models based on sensitive bands had a modeling R² range of 0.62-0.98 and a validation R² range of 0.36-0.94, demonstrating superior and more stable inversion results compared to other models. The study results indicate that establishing models to estimate soil carbon, nitrogen, and phosphorus content by controlling soil particle size is feasible, and selecting appropriate particle sizes can improve the accuracy of inversion models. Partial least squares regression (PLSR), as a high-precision inversion model, can help enhance the stability and predictive capability of the model, thereby more accurately estimating carbon, nitrogen, and phosphorus content in the soil. The research results provide solid theoretical and technical support for the quantitative inversion of surface soil carbon, nitrogen, and phosphorus in wetlands with different particle size treatments based on hyperspectral remote sensing.