Performance of soil total nitrogen monitoring model using optimal preprocessing combinations of in-situ and indoor spectra

To explore the performance of in-situ spectral monitoring of soil total nitrogen, this study focused on cotton fields with different soil textures and sampled soil from 0–60 cm depth. Three different preprocessing combinations were applied to the indoor and in-situ spectra, and four modeling methods including the Generalized Regression Neural Network (GRNN), Random Forest Regression (RFR), Support Vector Machine Regression (SVR), and Ordinary Least Squares Regression were optimized using the Northern Goshawk Optimization (NGO) algorithm. The goal was to establish and select the best model for monitoring soil total nitrogen content in each soil layer. The results showed that: (1) Compared with no preprocessing, different preprocessing combinations improved the model accuracy by 0.19–0.49. The optimal preprocessing combination for the surface soil was First Derivative (FD) - Standard Normal Variate (SNV) - Z-score - Savitzky-Golay (SG), and for the medium and deep soil was FD - SNV - Continuum Removal (CR) - SG. (2) The optimized NGO-GRNN model outperformed the GRNN model, with an improvement of 60%, 12%, and 19% in R ² for the shallow, medium, and deep soil layers, respectively. (3) The model constructed using indoor spectra performed better than the in-situ spectra in monitoring soil total nitrogen content. However, the in-situ spectra-based models for different soil layers had an R ² greater than 0.6, indicating good monitoring performance and eliminating the laborious steps of indoor spectral processing. This study provides theoretical and technical support for rapid acquisition of nutrient information in various soil layers of cotton fields using in-situ spectral monitoring, demonstrating feasibility and robustness.