Inversion of Leaf Water Content of Cinnamomum camphora Based on Preferred Spectral Index and Machine Learning Algorithm

Abstract

Plant leaf water content significantly influences photosynthetic efficiency and crop yield. Leaf water content (LWC) and equivalent water thickness (EWT) are indicators that reflect the water state within plant tissues, and they play a crucial role in assessing plant water supply and usage. In recent years, there has been a growing focus on the rapid and precise determination of plant water content. In this study, Cinnamomum camphora (C. camphora) was chosen as the subject of investigation. After acquiring spectral data, three types of vegetation indices were computed: the empirical vegetation index, the random combination dual-band vegetation index, and the ‘trilateral’ parameter. Four groups of optimal spectral index screening strategies were established, namely an empirical vegetation index group (G1), a random combination dual-band vegetation index group (G2), a ‘trilateral’ parameter group (G3), and a mixed group (G4). Three algorithms, specifically random forest (RF), radial basis function neural network (RBFNN), and support vector machine (SVM), were employed for the estimation of leaf water content (LWC) and equivalent water thickness (EWT) in mature C. camphora. The results demonstrated that the G4 group displayed superior performance, yielding five optimal spectral indices for LWC: water index (WI), optimized soil-adjusted vegetation index (OSAVI), difference vegetation index (DVI) at wavelengths 734 and 956 nm, first-order difference vegetation index (DVI-FD) at wavelengths 1009 and 774 nm, and red-edge amplitude (Dr). With regard to EWT estimation, the five optimal spectral indices encompassed the red-edge normalized difference vegetation index (RE-NDVI), simple ratio water index (SRWI), difference vegetation index (DVI) at wavelengths 700 and 1167 nm, first-order difference vegetation index (DVI-FD) at wavelengths 1182 and 1514 nm, and red-edge area (SDr). Utilizing these indices as inputs significantly enhanced the accuracy of the models, with the RF model emerging as the most effective for estimating LWC and EWT in C. camphora. Based on the LWC estimation model of the G4 group and the RF algorithm, the determination coefficient (R2) for both the training and test sets reached 0.848 and 0.871, respectively. The root mean square error (RMSE) was 0.568% for the training set and 0.582% for the test set, while the average relative error (MRE) stood at 0.806% and 0.642%, respectively. Regarding the EWT estimation model, R2 values of 0.887 and 0.919 were achieved for the training and test sets, accompanied by RMSE values of 0.6 × 10−3 g·cm−2 and 0.7 × 10−3 g·cm−2, and MRE values of 3.198% and 2.901%, respectively. These findings lay a solid foundation for hyperspectral moisture monitoring in C. camphora and offer valuable reference for the rapid assessment of crop growth status.