Aprendizado de máquina em dados de espectroscopia do infravermelho próximo para estimativa do teor de carbono orgânico total no solo

Abstract

With the advent of precision agriculture, new technologies have been adopted in the field. In recent years, theoretical frameworks and innovative practices have expanded, contributing to more sustainable and eco-friendly agriculture by offering solutions for processes previously deemed environmentally harmful and financially burdensome. Thus, the use of spectroscopy techniques—particularly NIRS—combined with machine learning methods has emerged as an alternative to mitigate pollution and reduce operational costs. This study aimed to develop predictive models for estimating soil organic carbon content using machine learning techniques. Specific objectives include building spectral data sets in the infrared region, identifying suitable machine learning techniques for the created datasets, and evaluating the potential of the generated models using the R² coefficient. Soil samples for this study were taken from a long term experiment lasting just over 30 years in Ponta Grossa, Parana State, Brazil. Six algorithms were employed in this study: Huber Regressor, Extra Trees Regressor, Catboost Regressor, ElasticNet, Lasso, and Bayesian Ridge Regressor. The Huber Regressor algorithm stood out as the best, considering the quadratic basis (R² = 0.86) and standard basis without transformation (R2 = 0.66). For the logarithmic basis, the greatest highlight was the Extra Trees Regressor algorithm (R² = 0.54).