Análise de espectroscopia utilizando aprendizado de máquina para quantificação de carbono e nitrogênio

Abstract

Near-Infrared (NIR) spectroscopy emerges as a promising alternative to traditional soil analysis methods, offering advantages such as speed, low cost, and non-destructibility. This work proposes the application of machine learning techniques to calibrate predictive models for carbon (C) and nitrogen (N) contents in Latosol and Cambisol soils, using NIR spectral data acquired with a portable equipment (MyNIR). Various pre-processing methods were tested, with the most promising being the Savitzky-Golay filter and outlier removal via the Nonlinear Iterative Partial Least Squares (NIPALS) algorithm with a Huber loss function. The data strategies were compared using 10-Fold Cross-Validation, Leave-One-Out, and Holdout via the Kennard-Stone method, followed by standardization. Stacking Ensemble Learning models were employed, using Partial Least Squares (PLS), Support Vector Machines Regression (SVR), and Ridge as base models, and Linear Regression as the meta-model. The models were trained and evaluated using the metrics 𝑅2, RMSE, MAE, and RPD. The final results indicated superior performance for Latosol (𝑅2 = 0.91 for C and 𝑅2 = 0.89 for N), while Cambisol presented satisfactory, though less expressive, results (𝑅2 = 0.79 for C and 𝑅2 = 0.77 for N). The difference between soil types suggests the influence of pedological characteristics. The models demonstrated RPD > 2.0 and low overfitting, validating the potential of the approach for rapid quantification of C and N. This study contributes to the optimization of sustainable agricultural practices, aligning with the demand for efficient and environmentally friendly analytical methods. The technique developed in this study enables faster decision-making for producers and consultants based on organic matter content, fertility indicators, and nutrient availability.