How Well Can Reflectance Spectroscopy Allocate Samples to Soil Fertility Classes?

Fertilization decisions depend on the measurement of a large set soil fertility indicators, usually through laboratory determination, which is costly and time-consuming. Visible near-infrared (vis-NIR) spectroscopy combined with machine learning can simultaneously predict various indicators. Spectroscopy inherently less accurate than direct determination. However, in many fertilization recommendation contexts, farmers mainly fertilize according to classified rather by continuous property values. These classes have defined limits We hypothesized that additional inaccuracy from may not be important for properties grouped into classes. This study compared indirect prediction Indirectly, (1) using vis-NIR spectra 20 indicators (pH, organic matter (SOM), cation exchange capacity (CEC), total nitrogen (TN), phosphorus (TP), potassium (TK), alkali-hydrolyzable (AN), available (AP), (AK), calcium (Ca), magnesium (Mg), silicon (Si), sulfur (S), boron (B), iron (Fe), manganese (Mn), copper (Cu), Zinc (Zn), molybdenum (Mo) chlorine (Cl)) (2) allocating Directly, predicting directly learning. The accuracy these two methods were accuracies needed acceptable class allocation determined. example dataset spectral library Guizhou Province, southwest China. model performance was evaluated overall tau index, accounts imbalance. For based three (low, medium high), eight (CEC, Cu, Si, Zn, S, Mn, Ca Mg), nine (B, AN, TK, AK, SOM, TN, TP, Fe Mo) (Cl, AP pH) within range 0.80–1.0, 0.60–0.80 0.40–0.60, respectively. same classes, (TN, CEC, pH, Fe, Mo SOM) (Cl AP) conclude fairly successful most properties, either or models. advantage models both specific values obtained at no increase cost, while are suggested when only information available.