CALIBRATION ALGORITHMS ASSESSMENT FOR SOIL NITROGEN PREDICTION WITH NEAR-INFRARED SPECTROSCOPY AND DATA AUGMENTATION

Abstract

Spectroscopy and machine learning are crucial in smart farming, enhancing soil variability management through predictive spectral models. Choosing suitable regression algorithms is essential due to complex soil-reflection relationships. Additionally, algorithms require a large amount of data to reach good performance, which can be challenging for researchers. Through specific metrics such as R², root mean square error, and residual predictive deviation (RPD), this study evaluates four regression algorithms for soil nitrogen prediction: Partial Least Squares (PLS), Extreme Learning Machine (ELM), Support Vector Machine (SVM), and Random Forest (RF). Models were built using near-infrared (NIR) spectroscopy and artificial data augmentation through generative adversarial networks. Spectral preprocessing was performed using a moving average smoothing and Savitzky-Golay derivative filter. The selection of spectral variables was carried out using a genetic algorithm. Artificial data augmentation improved model performance, with SVM and RF outperforming PLS and ELM, achieving RPD > 2, R² > 0.8, and lower error rates.