The integration of species information and soil properties for hyperspectral estimation of leaf biochemical parameters in mangrove forest

To understand the drivers of the mangrove degradation, more attention should be paid to the spatio-temporal dynamics of leaf biochemical parameters in response to changes of climatic or eco-environmental conditions in mangrove forests. With leaf reflectance of 135 mangrove samples collected from five sites, this study aimed to employ continuous wavelet transform (CWT, 128 scales) to estimate leaf biochemical parameters (SPAD-502 value, water content, nitrogen concentration, and phosphorus concentration) using random forest regression (RFR) model, and further to explore the effect of species and soil properties on the estimation of leaf biochemical parameters. Three dataset splitting methods (random splitting, RS; Kennard-Stone, KS; sample set partitioning based on joint x-y distance, SPXY) were employed to divide the original dataset into training (60% samples) and test set (40% samples). Compared with the 128 RFR models using original and first derivative reflectance (OR and FDR) spectra, the results showed that the models using wavelet power spectra (OR_CWT and FDR_CWT) at specific scales (<= 32) achieved better accuracy in estimating leaf biochemical parameters. The RFR models with SPXY method outperformed those with RS and KS methods, and leaf SPAD-502 and nitrogen estimation accuracies were higher than leaf water and phosphorus estimation. Compared with the RFR models using wavelet power spectra alone in leaf biochemical parameter estimation, the models using wavelet power spectra with the integration of species information and soil properties increased mean R-CV(2) (determination coefficient of cross-validation) values by 0.34%-23.73%, mean R-Val(2) (determination coefficient of independent validation) values by 0.24%-8.16%, and mean RPD (residual prediction deviation) values by 0.90%-11.95%. Moreover, the factors of species and soil total carbon showed major contribution to the RFR models in leaf SPAD-502 and N estimation. We concluded that the integration of species information and soil properties with SPXY and CWT method had great potentials in the accurate estimation of leaf biochemical parameters in mangrove forests, which could further help to understand the interaction between soil and mangrove plants, and to provide theoretical support for the ecological conservation and management of mangroves.