Vegetation aboveground biomass (AGB) is an important ecological parameter reflecting the functional status and health level of wetland ecosystems, which can provide key support for ecosystem stability analysis and carbon sink capacity assessment. In this paper, we take typical Phragmites australis (P.australis) wetlands in Binhai as the research object, and based on the high-resolution multispectral images acquired by UAV and 113 measured P.australis AGB samples, we systematically evaluate the modelling effect of different combinations of feature variables, and screen the variables by using the importance analysis of the Random Forest (RF) algorithm. Subsequently, three machine learning algorithm models, namely RF, Support Vector Machine (SVM) and Extreme Gradient Boosting (XGBoost), were constructed respectively to model and compare the accuracy of the screened feature variables, aiming at exploring feasible paths to carry out wetland vegetation biomass inversion based on low-altitude remote sensing technology, and establishing a high-efficiency estimation framework applicable to wetland ecological monitoring. The results show that the joint use of the variable combination of multispectral band reflectance and vegetation index is significantly better than the single variable type, which can effectively improve the accuracy of AGB estimation. Among the three modelling algorithms, the XGBoost model has the best performance, with a R2 of 0.731, a RMSE of 0.184 kg/m2, and a RPD of 2.431, which shows strong stability and robustness. The results of the spatial distribution analysis showed that the P.australis AGB in the study area had significant spatial heterogeneity, with the high value areas mainly distributed in the shallow water zone around the water body and in the low-lying areas, while the AGB values in the higher terrain areas were relatively low. Quantitative statistics showed that the total AGB of P.australis in the study area was 3038.78t, and the range of P.australis AGB values was 0.486—1.705kg/m2, with an average of 0.88kg/m2. Variable contribution analysis showed that the red edge band and its derived vegetation indexes, especially Modified Chlorophyll Absorption Reflectance Vegetation Index (MCARI) contributed more in the P.australis AGB inversion model, indicating its strong sensitivity in capturing changes in chlorophyll content and canopy structure of wetland vegetation, which significantly enhanced the estimation accuracy of AGB. This study not only verifies the feasibility and effectiveness of P.australis AGB inversion based on UAV multispectral images combined with machine learning algorithms, but also provides a scientific basis for subsequent wetland ecosystem function evaluation, carbon stock estimation and critical habitat monitoring.