Uso de algoritmos de aprendizado de máquina para a estimativa de propriedades químicas em amostras de solos

Abstract

Soil analysis is fundamental for agriculture, enabling fertility determination and proper fertilizer recommendation. However, traditional laboratory methods (wet chemistry) are costly, timeconsuming, and generate toxic chemical residues. Near-Infrared Spectroscopy (NIR) emerges as a fast, non-destructive, and sustainable alternative but faces challenges due to the complexity and overlapping of spectral signals. This work aimed to develop and evaluate machine learning models to estimate the concentration of chemical elements (such as Magnesium, Potassium, Phosphorus, Calcium, Organic Matter, and pH) in Brazilian soil samples using NIR spectral data. The methodology involved spectral pre-processing with techniques such as SNV, MSC, and derivatives (Savitzky-Golay), followed by the application and optimization of regression algorithms: PLSR, SVR, RF, XGBoost, andStacking. The results indicated an overall modest performance, reflecting the high heterogeneity of the samples. Tree-based and ensemble models (such as XGBoost, RF, andStacking) outperformed classical linear methods (PLSR) in most cases, managing to partially capture the variability of attributes like Phosphorus and Magnesium. It is concluded that, although the technique is promising, soil variability requires more robust calibration strategies for effective practical application.