Collaboratively classifying remote sensing imagery for Land Cover maps of Aruba

Abstract

Land cover changes are vital in shaping ecosystems, impacting food production, carbon sequestration, and essential services. Human-driven land cover alterations rapidly occur worldwide, leading to significant environmental and socio-economic challenges. This thesis focuses on creating land cover maps for Aruba, a small island nation facing environmental pressures from rapid urbanization. Accurate land cover maps can assist policymakers in addressing these challenges by providing essential data for sustainable land management. This study tackles two main challenges to developing these maps: selecting suitable land cover classes and determining the best machine learning classifiers.

The land cover classes were defined using the WorldCover project as a base, and a hierarchical set of land cover classes was created using input from the Aruban Department of Nature and Environment (DNM) to ensure relevance to local conditions. Multiple machine learning classifiers were tested to determine the most accurate methods for classifying Sentinel-2 imagery into these hierarchical land cover classes. The final land cover maps were then used to fill critical knowledge gaps identified by the DNM. These gaps include the need for data on environmental indicators (SDG 15.1.1: forest cover, SDG: 15.3.1 degraded areas, and BGF: A.2 natural area) and the Build with Nature policy, which integrates conservation with infrastructure planning.

The results of this study showed that K-Nearest Neighbours (KNN) was the best-performing classifier for main land cover classes, achieving an accuracy of 70.49%. These land cover maps with the best-performing classifiers are accessible via a Google Earth Engine application, continually supplying information for future policy development and environmental management on the island. The indicator values for the year 2024 were found to be 2.28% for SDG 15.1.1, 16.24% for SDG 15.3.1, and 70.57% for BGF A.2. Furthermore, the land cover maps provided valuable insights for the Build with Nature policy, offering spatial data that supports sustainable infrastructure planning.