Abstract: The rapid degradation of wetland ecosystems threatens the stability of biodiversity, thus requiring efficient global monitoring and conservation technologies to protect wetlands. Compared with traditional field-based sampling, remote sensing provides substantial advantages for estimating wetland plant diversity and is grounded in several theoretical hypotheses. Among these, the spectral variation hypothesis (SVH) states that spectral reflectance is jointly influenced by environmental heterogeneity and plant diversity, thereby establishing a correlation between the two. This hypothesis is simultaneously affected by spatial scale effects (e.g., pixel size, grain size, plot size) and background effects. Among these, preprocessing to remove environmental background interference is a critical step for reducing environmental noise. At present, the mainstream background-removal approaches include the visual threshold method (VTM), spectral unmixing, and the threshold value optimization (TVO) method. Both the VTM and TVO remove background pixels using vegetation-index (VI) thresholds, while the TVO provides a more objective criterion for threshold selection. However, TVO has mainly been applied in homogeneous environments containing only soil and vegetation, and its performance in complex natural wetlands remains insufficiently validated. To address this gap, this study applied both the TVO and the VTM to unmanned aerial vehicle (UAV) imagery collected in the Sanjiang National Nature Reserve, aiming to evaluate how background-removal strategies influence the capability of vegetation indices (VIs) to estimate species diversity and to determine the applicability of the two approaches. The results show that, compared with the VTM, the TVO generally improved the goodness of fitting of VIs means. In particular, NDREI (Normalized Difference Red-Edge Index) and MTCI (MERIS Terrestrial Chlorophyll Index) exhibited the highest adjusted coefficients of determination (R²_adj) with species richness (R²_adj=0.35-0.56 and 0.42-0.54, respectively). In contrast, improvements in VIs standard deviations were limited, with near-infrared-based indices showing relatively strong relationships only with the Shannon-Wiener index (R²_adj=0.00-0.34). VIs derived from identical spectral bands produced similar R²_adj patterns, whereas the standard deviation of near-infrared-related indices more effectively captured spatial heterogeneity. Under varying background interference, differences in R²_adj between the Shannon-Wiener and Simpson indices may be attributed to their differing sensitivities to rare species and species evenness. Results from the two background-removal approaches further indicate that the TVO is more objective and suitable for ecologically homogeneous environments, whereas the VTM, despite its subjectivity, is more appropriate for complex ecological environments. Therefore, combining both approaches is recommended to remove environmental background effects and fully exploit their complementary advantages.