Artificial intelligence with wrapper multioutput regression to identify chemical contents of fresh and dried manure samples using NIR spectroscopy.

Analyzing manure nutrients such as total ammonium nitrogen (NH), dry matter (DM), calcium oxide (CaO), total nitrogen (-N), phosphorus pentoxide (PO), magnesium oxide (MgO), and potassium oxide (KO) helps in fulfilling crop nutritional needs while improving the profitability and a lower risk of pollutants. This study used two Near Infra Red (NIR) spectral datasets of fresh and dried manure. The freshly prepared NHCl, CaO, Ca(OH), PO, MgO, and KO samples were used for spectral signature peak identification and calibration. The pre-processing methods, Savitzky-Golay (SG), Multiplicative Scatter Correction (MSC) and Standard Normal Variate (SNV), were applied to correct the spectra and PCA to reduce the dimension of these pre-processed spectra. Several regressors and classifiers with wrapper multioutput regression algorithms (WMRA) and stacking generalization models were applied to identify the chemical contents. In the regression analysis, the WMRA achieved the highest coefficient of determination (R) of 0.983 and RPD of 7.68 for fresh manure, while R of 0.949 and RPD of 4.42 for dried manure. In classification analysis, k-nearest neighbors (KNN) achieved 100.0% highest accuracy, while 98.1% 5-fold average accuracy. The overall findings of this study indicate that NIR, in conjunction with WMRA and KNN, has a high potential for quick detection of nutrients in manure.