Coupling critical nitrogen dilution curve with hyperspectral feature optimization and machine learning for inversion of nitrogen nutrition index in greenhouse cucumber.

Introduction: Rapid, non-destructive, and accurate diagnosis of nitrogen (N) nutritional status in greenhouse cucumber production is critical to mitigate over-fertilization risks and support sustainable intensification. Methods: An integrated diagnostic framework was developed by coupling the agronomic critical nitrogen (Nc) dilution curve with a hybrid Competitive Adaptive Reweighted Sampling and Variance Inflation Factor (CARS-VIF) feature selection algorithm, followed by benchmarking six machine learning regression models, including K-Nearest Neighbors (KNN), Multilayer Perceptron (MLP), Support Vector Regression (SVR), Extra Trees, Random Forest (RF), and Gradient Boosting Regression (GBR) . Results: The critical nitrogen dilution curve was parameterized as Nc = 4.378 × DW-0.115 (R² = 0.745), with empirical Nitrogen Nutrition Index (NNI) values ranging from 0.72 to 1.22. The CARS-VIF pipeline compressed the full spectrum down to four sensitive wavebands (430, 677, 688, and 953 nm), achieving a 99.81% dimensionality reduction. GBR emerged as the optimal inversion model, delivering a validation R² of 0.845, RMSE of 0.061, and the lowest MAE of 0.046. Discussion: This non-destructive framework holds promise for future integration into automated greenhouse sensor platforms or unmanned aerial vehicles (UAVs), offering a powerful digital toolkit to support precision fertilization decision-making and sustainable smart facility horticulture. Additional validation under field conditions is required before operational deployment.