Quantitative analysis of nitrogen and phosphorus in dairy cow slurry based on near-infrared spectroscopy combined...

Rapid and accurate detection of total nitrogen (TN) and total phosphorus (TP) in dairy cow slurry is essential for guiding the scientific application of slurry to fields, preventing soil nutrient imbalance and water pollution, and promoting sustainable agricultural development and resource utilization of slurry. However, the complex composition and high moisture content of slurry, along with the high dimensionality and nonlinear characteristics of its spectral data, pose challenges to accurate and efficient prediction. In this study, near-infrared spectroscopy was combined with partial least squares regression (PLSR), extreme gradient boosting (XGBoost), and deep neural network (DNN) to establish TN and TP prediction models for dairy cow slurry using different spectral preprocessing methods (MSC, SG, and MSC + SG). The optimal preprocessing methods for TN and TP were determined according to model performance. Furthermore, uninformative variable elimination (UVE) and bootstrap soft shrinkage (BOSS) were employed to extract characteristic wavelength variables and construct models, and the performances of different models were compared to identify the optimal one. Results showed that the DNN model achieved the best performance in TN and TP prediction, with R_p² values of 0.92 and 0.75, and RPD values of 3.49 and 2.01, respectively, significantly outperforming PLSR and XGBoost. When using UVE and BOSS, both methods yielded comparable predictive performance; however, BOSS achieved a higher degree of data compression and modeling efficiency, showing greater potential for practical applications. Overall, the BOSS-DNN approach demonstrated remarkable advantages in predicting complex components of dairy slurry and provides a new technical pathway and theoretical foundation for rapid on-site nutrient detection of livestock and poultry slurry.