Abstract: Mangrove forest is a natural coastal defense barrier, which plays an irreplaceable role in coastal disaster prevention and mitigation. Therefore, it is important to understand the growth status of mangroves. The ecological stoichiometry of plants can reflect their nutrient storage and supply capacity. Applying hyperspectral data to quantify the ecological stoichiometry of mangrove plants and exploring the accuracy and stability of hyperspectral retrieval of leaves can provide a technical reference for rapid remote sensing monitoring of mangrove growth conditions. In this study, we collected hyperspectral data of leaves of three dominant mangrove species (Bruguiera sexangula, Ceriops tagal, and Rhizophora apiculata) in Qinglangang Mangrove Nature Reserve, Hainan, and retrieved the contents and stoichiometry of C, N, and P. The results showed that there were significant differences in the contents and stoichiometry of C, N, and P among the three species, indicating differences in nutrient utilization of the three mangrove species. The Random Forest (RF) model outperformed Back Propagation Neural Network (BPNN) model in retrieving C, N, P contents and their ecological stoichiometry considering R², RMSE and RPD. This study demonstrated that the contents and stoichiometry of C, N, and P in mangrove leaves could be accurately estimated by leaf hyperspectral data. RF model is recommended for the hyperspectral retrieval of mangrove ecological stoichiometry when considering model accuracy and robustness.

Key words: machine learning models, hyperspectrum, ecological stoichiometry, mangrove