ESPECTROSCOPIA VIS-NIR-SWIR, INDICADORES MICROBIANOS E CLASSIFICAÇÃO DO USO DO SOLO UTILIZANDO MODELOS DE APRENDIZADO DE MÁQUINA

Abstract

Soil microbiology demands analytical methods and non-destructive techniques in laboratory analyses that are capable of evaluating soil quality and its microbiological markers. Therefore, spectral sensing, which is widely used in agriculture as an effective and precise tool for agricultural diagnosis, associated with machine learning techniques, may present potential to perform classifications through models that use specialized algorithms. Thus, the objective of this research was to classify different types of land use through microbial indicators and spectral sensing using machine learning models. The study was conducted at Rancho Ideal farm, located in the municipality of Chapadão do Sul, Mato Grosso do Sul. Soil samples were collected from four different land uses present in the farm plots, namely: Native Forest (Cerrado), Eucalyptus, Pasture and Cropland. In each land use, 25 samples were collected, totaling 100 samples. Spectral acquisition was performed using the Ocean Optics equipment consisting of two sensors: STS-VIS and Flame. The microbial indicators of soil quality analyzed were: microbial biomass carbon, basal soil respiration and metabolic quotient, through specific methodologies. For modeling, six models were trained globally with the complete dataset for each indicator, and the evaluation was performed through cross-validation with ten repetitions. The results show that, using the Savitzky-Golay preprocessing technique, the tested models were capable of identifying complex patterns and correlations between the databases used, especially those based on gradient boosting, reaching an average accuracy of 94%. Therefore, it is concluded that for classification, in a consistent spectral pattern, models of different natures obtained significant performance, without depending on a specific algorithm.