Espectroscopia Vis-NIR-SWIR, MIR e fluorescência de raio-X na determinação de nitrogênio, carbono orgânico total e frações húmicas em Latossolo

Abstract

The significant growth in future demand for food highlights the importance of increasing productivity in agricultural crops and sustainable planning to preserve the environment and produce in a way that mitigates environmental variations. The vitality of a soil in an agricultural system depends on carbon transformations, nutrient cycling and good structure. The crop intercropping system stands out as a sustainable practice for agricultural production, offering crop diversification and restructuring soil fertility. Currently, the concept of soil health values the reduction in the application of chemical products, conservation of microbial biodiversity, maintenance of organic matter, among other principles. Thus, with the advancement of technology and the dissemination of practical and precise knowledge in agriculture, the assessment of soil fertility has adapted to new tools, capable of complementing or even replacing existing analysis methods aimed at predicting nutrients in the soil. Due to its high efficiency, hyperspectral sensing in conjunction with machine learning can evaluate soil properties, such as organic matter, clay, pH, CEC and nutrients present in the soil. Spectroscopic techniques covering the Vis-NIR-SWIR, MIR and X-ray Fluorescence wavelengths can quickly provide estimates of soil nutrient content at a reduced cost. The general objective of this agronomic research was to study calibration and model validation methods for quantifying nitrogen, total organic carbon and humic fraction in a medium sandy Oxisol via spectroscopy. Chapter 1: The soil used was collected in an experimental area located on the experimental farm of the Universidade do Oeste Paulista (UNOESTE) geographic coordinates, South latitude 22°17’13’’ and West longitude 51°40’34’’. The experiment was then conducted in an air-conditioned greenhouse at the University. The experimental design was completely randomized with four doses of the ammonium nitrate compound (0, 100, 200 and 400 kg ha-1) and two collection depths 0-20 and 80-100cm, containing ten replications. Spectral analyzes were carried out using Vis-NIR-SWIR, MIR and XRF spectroscopic techniques, and the application of machine learning algorithms Cubist, PLSR, Random Forest and Support Vector Machine for the development of predictive models. Chapter 2: The soil used was collected in an experimental area located on the experimental farm of the Universidade do Oeste Paulista (UNOESTE). The experiment has a randomized block design with four treatments: 1) Mombasa (Megathyrsus maximus cv. Mombasa) with nitrogen fertilization, 2) Mombasa intercropped with Guandu Beans (Cajanus cajan), 3) Mombasa intercropped with Macrotyloma axillare cv. Java, 4) Mombasa without nitrogen fertilizer. The models that presented the best predictive performance were validated.