Artificial intelligence-assisted prediction of nitrogen, phosphorus, and potassium contents in ecoenzymes using FT-NIR spectroscopy

ABSTRACT The main challenge lies in the instability of spectral data due to variations in raw materials and fermentation conditions, which often leads to decreased prediction accuracy. This study aims to develop a rapid prediction model for nitrogen (N), phosphorus (P), and potassium (K) content in ecoenzymes using FT-NIR spectral descriptors (800–2500 nm) integrated with the automatic combination of preprocessing and hyperparameter (ACPH) method and ensemble machine learning (EML). A total of 70 NIR spectra of ecoenzyme samples were analyzed with 1001 spectral features and divided into calibration and prediction data with a 90:10 ratio using 5-fold cross-validation. The PLSR-based ACPH model, optimized with 12 types of preprocessing and 9 latent variables, was compared with the EML model, which combines 11 basic algorithms through a PLSR meta-learner. The results showed that the EML model provided the best performance for predicting N (Rp 2 =0.958; RMSEP=0.0013%), P (Rp 2 =0.809; RMSEP=0.1118%), and K (Rp 2 =0.822; RMSEP=0.0351%) content, with RPD values between 2.285–4.872 indicating high reliability. The integration of FT-NIR descriptors with ensemble learning algorithms demonstrates the great potential of applying artificial intelligence for rapid monitoring of agricultural waste-based ecoenzyme macronutrients, thereby supporting environmentally friendly production and efficiency in organic resource management.