This work proposes a methodology for generating a high-quality dataset of samples to train machine learning models for land use and land cover (LULC) classification, representing an initial step toward the production of large-scale LULC maps. The novel approach is designed to generate a robust set of multi-pixel, multi-temporal, and multi-band samples by integrating label generation and remote sensing data processing across two main phases. In the first phase, multidimensional data are collected from remote sensing sources, while label information is derived by comparing harmonized LULC maps, along with additional processing to enhance data quality. In the second phase, samples are extracted to create a diverse, representative, and balanced set of reliably labeled multidimensional data for LULC classification. In this study, reference maps were obtained from the TerraClass and MapBiomas projects, and satellite imagery was sourced from the Brazil Data Cube platform.