Spectral Estimation of Carotenoid Density in Populus pruinosa Leaves

Carotenoids play a crucial role in the photosynthesis process in plants. Estimating and modeling the carotenoid content in Populus pruinosa leaves via high-spectrum technology is highly important for health status monitoring. This study involved acquiring the spectral reflectance of Populus pruinosa leaves at different times, followed by smoothing the data with a Savitzky—Golay filter, and then using methods such as first derivative (FD), continuous wavelet transform (CWT), and first-order derivative combined with continuous wavelet transform (CWT+FD), creating three spectral transformation methods. Two- and three-dimensional vegetation indices were then constructed in a unified manner. Two modeling methods, backpropagation neural network (BPNN) and support vector regression (SVR), were employed to estimate the leaf carotenoid density by combining the vegetation indices. The results show that after the spectral reflectance of the canopy of Populus pruinosa is processed by FD, CWT, and CWT+FD on the basis of SG smoothing, it can effectively highlight the spectral characteristics of Populus pruinosa leaves, and the local spectral absorption features are more significant. Compared with the three spectral preprocessing methods, the results showed that the correlation between the values processed by the FD + CWT method and the leaf carotenoid density is the highest. The constructed three-band vegetation index exhibited a 4.26% stronger correlation with carotenoid density than did the two-band vegetation index. Among the three-band index-based models, the SVR model outperforms the BPNN model. For chlorophyll density, the SVR model based on the three-band index processed using CWT+FD achieves the best performance. The coefficient of determination (R2) for the SVR model set was 0.782, the root-mean-square error (RMSE) was 0.022, and the relative percentage deviation (RPD) was 0.206. For the validation set, the (R2) value was 0.648, the RMSE was 0.023, and the RPD was 1.526, indicating the best model accuracy.