Accurate and objective measurements of pasture biomass is a fundamental step in improving productivity in a grazing system, thus allowing cattle breeders to better meet their animals' feeding needs. Therefore, new methodological approaches that characterize the horizontal structure of pasture should be investigated. Due to the synoptic, multispectral and revisit characteristics of the satellites, remote sensing images allow to obtain fundamental information for spatial and temporal characterization of pastures, as well as biomass estimation along the forage development using time series. However, the process of pasture biomass estimation using time series is not a simple process, it involves several factors, such as the interpreter's experience, knowledge of the study areas, temporal signature of the targets, as well as the methodology used. Currently, the most advanced methods of biomass estimation using remote sensing images use machine learning algorithms. However, remote sensing data generate a large number of predictor variables, which makes modeling biomass estimation a challenge. Thus, data mining techniques for feature selection can add improvements in the performance of this modeling. Therefore, this project aims to test feature selection methods and estimate pasture biomass in a crop-livestock integration system using Sentinel-2A time series and machine learning.
News published in Agência FAPESP Newsletter about the scholarship: