Estimation of Korla Fragrant Pear Leaf Water Content Using Near-Infrared Spectroscopy Combined with Machine Learning

In modern agricultural production, accurately estimating the leaf water content (LWC) of Korla fragrant pear is crucial for achieving scientific irrigation and ensuring fruit quality. However, constructing accurate and effective LWC prediction models remains challenging due to limitations in sample selection, spectral feature analysis, and model applicability. To address these issues, this study was conducted to systematically optimize the process. During sample collection, a random split method was employed to divide the dataset into modeling and testing sets at a ratio of 75%:25%. This approach ensures computational efficiency, avoids data leakage, and balances training and evaluation needs, particularly for small- to medium-sized datasets. Specifically, in stage S1, 352 samples were allocated to the modeling set and 108 to the testing set, while in stage S2, 137 and 58 samples were assigned, respectively. The analysis revealed slight differences in LWC distribution and standard deviation between the modeling and testing sets, validating the scientific rigor of dataset division. For instance, the LWC distribution in the S1 modeling set ranged from 4.88% to 83.45%, with a standard deviation of 11.33%. The spectral acquisition process within the range of 4000 cm-1 to 10,000 cm-1 exhibited complex absorbance variation trends, showing distinct characteristics across different intervals. Preprocessing techniques such as SG convolution smoothing, MSC, and SNV significantly reduced the absorbance variability and enhanced spectral features. Notably, the selection of LWC feature bands differed markedly between stages S1 and S2. For example, in S1, SNV-SPA (successive projections algorithm) feature bands were concentrated around 5000 cm-1, 6000 cm-1, and 7000 cm-1, whereas their positions shifted significantly in S2, reflecting the growth dynamics of the Korla fragrant pear. During the model-building phase, various algorithms, including Random Forest Regression (RFR), Backpropagation Neural Network (BP), and Support Vector Regression (SVR), were compared. Under different feature selections, the RFR model demonstrated strong predictive ability with determination coefficients (R2) exceeding 0.75 and root mean square errors (RMSE) below 0.7%. Specifically, the SNV-CARS-BP model achieved an R2 of 0.81594 in S1, while the SNV-SPA-RFR model reached an R2 of 0.817756 in S2, with relative deviations between the predicted and actual values of less than 5%. These results provide robust support for the precise LWC monitoring of Korla fragrant pear and offer valuable insights for subsequent research.