Non-Destructive Detection of Nutritional Elements in Fresh Tea Leaves Using Hyperspectral Technology Combined with a Multi-Stage Feature Selection Strategy

This study addresses two key challenges in tea nutrient analysis: the limited range of detectable nutrient elements in tea gardens and the interference caused by moisture in fresh tea leaves during spectral data acquisition. To overcome these issues, hyperspectral technology combined with effective spectral intelligent processing algorithms was used to develop quantitative, non-destructive prediction models for four essential nutrients—nitrogen, phosphorus, potassium and carbon—in fresh tea leaves. This study utilises EPO to address the issue of moisture interference in the spectra of fresh tea leaves, and combines it with SG for spectral data processing, namely SG-EPO. Through comparative analysis with traditional pre-processing algorithms, this method was found to effectively reduce moisture interference in fresh tea leaves and enhance prediction accuracy (R2). Finally, based on a multi-stage feature selection strategy involving SG-EPO-VCPA-IRIV-SVM_RFE and SG-EPO-BOSS-SVM_RFE, and in combination with three machine learning models—XGBoost, BP and SVR—quantitative prediction models were developed. The results indicated that the R2 for nitrogen is 0.896; phosphorus, 0.954; potassium, 0.913; and carbon, 0.928. The RMSEP values for N, P, K, and C were 0.062, 0.075, 0.438, and 0.157, respectively. Using this model, nitrogen, phosphorus, potassium and carbon contents were rapidly predicted in tea leaves after the exogenous application of GABA at different concentrations, enabling an assessment of the effects of exogenous GABA application on these nutrient levels. Furthermore, the Shapley Additive Explanation method was employed to identify the feature wavelengths that had the greatest contribution to the XGBoost model, effectively explaining the information underlying the improved model predictions regarding the correlation between spectral and chemical values. Finally, the accuracy of the predictions was confirmed using 20 samples from independent data, demonstrating that the proposed model can achieve rapid, non-destructive detection of multiple nutrient elements in fresh tea leaves under in situ conditions in tea plantations.