Prediction Model of Chlorophyll Content in Uncaria rhynchophylla Leaves Based on Machine Learning Algorithms

To address the need for rapid monitoring of the physiological status of the medicinal plant Uncaria rhynchophylla, hyperspectral technology was used to develop prediction models for leaf chlorophyll content (LCC). The effects of different combinations of sample partitioning methods, feature extraction techniques, and machine learning algorithms on model performance were compared to identify the optimal modeling strategy for estimating LCC in U. rhynchophylla. A total of 270 leaf samples were collected from the Liziping National Nature Reserve in Sichuan Province, China. Hyperspectral data covering 350–2500 nm were preprocessed using Savitzky–Golay smoothing. The samples were divided into calibration and prediction sets at a ratio of 8:2 using the Kennard–Stone (KS) algorithm and the sample set partitioning based on joint X–Y distances (SPXY) algorithm. Spectral features were extracted using the successive projections algorithm (SPA) and principal component analysis (PCA). Prediction models were subsequently developed using partial least squares regression (PLSR), a backpropagation neural network (BPNN), random forest regression (RFR), an extreme learning machine (ELM), and a support vector machine (SVM). Model parameters were optimized through grid search combined with five-fold cross-validation. Model performance was evaluated using the coefficient of determination (R²), root mean square error (RMSE), relative root mean square error (rRMSE), and adjusted coefficient of determination (adjusted R²). Correlations between the extracted features and LCC were also analyzed. Under KS and SPXY partitioning, SPA selected 7 and 12 characteristic wavelengths, respectively. PCA retained four principal components under both partitioning methods, with cumulative variance contributions of 96.82% and 96.54%, respectively. The prediction-set R² values of the models under KS partitioning ranged from 0.639 to 0.766, generally exceeding those obtained under SPXY partitioning, which ranged from 0.435 to 0.575. SVM achieved the highest predictive performance in the KS–SPA, KS–PCA, and SPXY–PCA combinations. Among all models, KS–PCA–SVM performed best, with a prediction-set R² of 0.766, an RMSE of 0.0092, and an rRMSE of 0.567%. Correlation analysis showed that the wavelengths most strongly correlated with LCC under KS–SPA and SPXY–SPA were 751 nm (r = 0.658, P<0.01) and 759 nm (r = 0.729, P<0.01), respectively, both located in the red-edge–near-infrared transition region. Under both PCA conditions, PC₃ exhibited the strongest positive correlation with LCC. The performance of hyperspectral models for predicting LCC in U. rhynchophylla was jointly affected by the sample partitioning method, feature representation approach, and machine learning algorithm. Under the conditions of the present dataset, KS–PCA–SVM demonstrated the best predictive ability. The spectral region near the red edge at 751–759 nm showed a relatively stable response to variations in LCC and may therefore serve as an important sensitive region for monitoring chlorophyll in U. rhynchophylla leaves. These findings provide a methodological basis for rapid and nondestructive monitoring of the physiological status of U. rhynchophylla leaves. Nevertheless, the model requires further validation across broader LCC gradients and under different seasonal and regional conditions.