Spectral prediction of anthocyanin concentration in Populus pruinosa leaves based on vegetation index

Populus pruinosa is the key foundation tree species in desert riparian forests in arid areas of northwestern China. Timely and accurate monitoring of the physiological status of P. pruinosa is crucial for restoring the damaged ecosystem. Anthocyanins are one of the important physiological indicators that reflect the environmental adaptability of P. pruinosa under stress. Existing studies have extensively applied hyperspectral technology for the quantitative prediction of crop leaf pigments. However, research on hyperspectral prediction of anthocyanin concentration in woody halophytes is still lacking, particularly in the integration of spectral preprocessing, species-specific vegetation index construction, and machine learning modeling.

In this study, the hyperspectral technology was used to estimate the anthocyanin concentration of P. pruinosa leaves collected in five months (June - October) under five groundwater depth conditions (0-2, 2-4, 4-6, 6-8, and 8-10 m). Based on first-order (FD) and second-order (SD) derivative processing, competitive adaptive reweighted sampling (CARS), https://xueshu.baidu.com/usercenter/paper/show?paperid=bea4d6371f19161f21aac22941cc4408&site=xueshu_se shuffled frog leaping algorithm (SFLA), and recursive feature elimination with cross-validation (RFECV) were used to extract spectral features of P. pruinosa leaves to construct the anthocyanin reflectance index, composite index, difference vegetation index, and normalized anthocyanin reflectance index. After that, the top 10 sets of data with high correlation with anthocyanin concentration were selected from each vegetation index to form a total data set (40 sets in total) for modeling. Twelve models were constructed using support vector machine (SVM) and one-dimensional convolutional neural network (1D-CNN) methods.

The FD and SD derivative transformations of the spectral reflectance significantly enhanced the correlation with anthocyanin concentration. The feature extraction methods SFLA and RFECV were superior in extracting the bands highly related to anthocyanin concentration, and the vegetation indices constructed based on these two methods had a high correlation with anthocyanin concentration in the red and near-infrared regions. The optimal prediction model was FD-SFLA-SVM (R² = 0.852, RMSE = 86.851 mg m^-2, RPD = 2.596).

Unlike existing vegetation index-based studies, the research develops a systematic approach to construct vegetation indices and models for estimating the anthocyanin concentration in the woody halophyte P. pruinose in deserts. The research will provide technical support for non-destructive monitoring of the physiological status of P. pruinosa, and also contribute to the restoration of desert riparian ecosystems.