Imaging spectroscopy has become a pivotal technique for estimating plant traits at the canopy scale. Accurate trait prediction is critical for biodiversity conservation, yet research on canopy traits in heterogeneous wetlands with complex species mixtures remains scarce. While the Community-Weighted Mean (CWM) method has been widely used for upscaling leaf traits to the canopy level, it often suffers from low model precision, and the suitability of alternative upscaling methods for predicting canopy mean traits using imaging spectroscopy remains uncertain. This study proposed a novel approach for calculating canopy mean traits using the geometric mean method and compared its performance to that of the CWM methods in combination with three modeling algorithms Partial Least Squares Regression (PLSR), Random Forest regression (RF), and Support Vector Machine regression (SVM). The accuracy was evaluated by exploring the predictive ability for nine canopy mean traits by using high spatial resolution UAV multispectral data. The analysis focuses on a wetland ecosystem characterized by high species diversity and hydrological variability, where precise plant trait estimation is essential for ecological process modeling. The results demonstrated that the geometric mean method yielded the highest validation accuracy for most canopy mean traits when paired with the SVM model (e.g., R² for N = 0.64, SLA = 0.38, and cellulose = 0.33). Notably, the geometric mean method, combined with UAV multispectral data, significantly enhanced the predictive performance for N, surpassing even that of hyperspectral data. This study underscores the potential of the geometric mean method for upscaling leaf traits to canopy traits. These findings contribute to advancing the prediction accuracy of plant functional traits through remote sensing techniques, while future studies may explore the integration of deep learning methods.