In remote sensing image analysis, reference data play a crucial role but are often inaccurate and uncertain. In this article we present a methodology for improving remote sensing reference data in three steps: segment realignment, evaluation via crowdsourcing and creation of a good quality dataset.