The RLCMS landcover map is developed through a rigorous workflow, starting with satellite imagery from Landsat missions (5, 7, 8, 9) and supplementary data sources such as the Joint Research Centre (JRC) and Shuttle Radar Topography Mission (SRTM). The raw satellite imagery undergoes various preprocessing steps like cloud masking, BRDF correction, and topographic correction. This ensures the creation of stabilized, high-quality datasets.
Next, relevant features are selected using covariates and reference data to train machine learning models and neural networks, which classify the land cover into basic building blocks known as "primitives." These primitives are then combined using a decision tree framework to create comprehensive land cover classes. The decision tree also allows for uncertainty assessment, providing a robust measure of the map's accuracy.
For a more detailed explanation of the process, please refer to the full Methodology Document.