Near-infrared spectroscopy and machine learning to estimate the physical and chemical properties of soils cultivated with Ilex paraguariensis

Abstract

Accurately estimating the chemical and physical properties of soil is crucial for monitoring nutrient dynamics in Ilex paraguariensis, aiming to enhance the efficiency and precision of fertilizer application. Nevertheless, the high costs associated with traditional analytical methods pose a significant barrier to their widespread use across numerous samples, underscoring the importance of developing more accessible techniques. Consequently, this study aimed to assess the efficacy of various combinations of multivariate methods and preprocessing techniques based on near-infrared spectroscopy in estimating the chemical and physical properties of soils cultivated with I. paraguariensis across five regions of Rio Grande do Sul (southern Brazil). These samples originated from the state’s five main yerba mate-growing regions: Região dos Vales (n = 6), Palmeira das Missões (n = 14), Alto Uruguai (n = 19), Nordeste Gaúcho (n = 19), and Alto Taquari (n = 49), totaling 107 samples. The analyzed chemical properties included pH in water, soil organic matter, available concentration of seven nutrients (phosphorus, potassium, sulfur, copper, zinc, boron, and manganese), and exchangeable concentration of aluminum, calcium, and magnesium). Clay content was evaluated as physical property. NIR spectra (780–2500 nm) were acquired for all soil samples. The multivariate learning models tested included partial least squares regression (PLSR) and support vector machines (SVM), combined with three spectral preprocessing techniques: detrending (DET), Savitzky-Golay derivative (SGD), and standard normal variate (SNV), with raw spectra (RAW) as the control. Model performance and preprocessing technique effectiveness were assessed using the coefficient of determination, root mean square error, and the ratio of performance to interquartile distance. The SVM showed superior predictive performance compared to PLSR, with preprocessing techniques improving estimation accuracy in the following order: RAW<SNV<DET<SGD. The most effective results were achieved by combining SVM models with SGD preprocessing. These findings indicate that selecting the appropriate mathematical models significantly enhances prediction of physical and chemical soil properties in samples cultivated with I. paraguariensis.