Machine Learning–Based Estimation of Leaf Nitrogen Content in Greenhouse Cucumber Using Spectral Data and SPAD Measurements

Accurate and non-destructive diagnosis of leaf nitrogen content (LNC) is critical for improving nitrogen use efficiency in greenhouse cucumber production. However, strong physiological variation across growth stages limits the reliability of single-sensor approaches. In this study, leaf spectral reflectance and soil plant analysis development (SPAD) measurements were collected under four nitrogen levels (0, 135, 270, and 540 kg·ha−1) at early, mid-, and late fruiting stages. Multiple machine learning models were developed using raw spectral bands (SP), vegetation indices (VIs), and SPAD data, and evaluated using the coefficient of determination (R2), root mean square error (RMSE), and relative error (RE). Results showed that SPAD showed a significant positive correlation with LNC across all stages, with the strongest relationship observed at the mid-fruiting stage (R = 0.7975). Model performance exhibited clear stage dependence. Using single features, the best R2 reached 0.800 (SP, early stage) and 0.794 (VI, early stage), but declined substantially at later stages. In contrast, integrating SPAD with spectral features significantly improved prediction accuracy, particularly at mid- and late stages. For example, the RF model based on SP + SPAD achieved R2 values of 0.917 and 0.901 at the mid- and late fruiting stages, respectively, with low RMSE and RE. Similarly, the VI + SPAD combination achieved R2 up to 0.893 at the late stage. Moreover, optimal algorithms varied across growth stages: SVR performed best at the early stage (R2 = 0.819), RF at the mid stage (R2 = 0.889), and XGBoost at the late stage (R2 = 0.842) under full feature fusion. These results demonstrate that model accuracy is jointly regulated by growth stage, feature composition, and algorithm selection. Overall, this study highlights that a growth-stage-specific data fusion strategy integrating SPAD and spectral features is essential for robust LNC estimation, providing a practical framework for precision nitrogen management in greenhouse cucumber production.