Accuracy assessment of global land-cover datasets for monitoring forest landscape restoration targets in Malawi

Malawi contributes to the global restoration movement through its commitments under the UN Decade on Ecosystem Restoration and the Global Biodiversity Framework. With deforestation rates among the highest in sub-Saharan African countries, Malawi has pledged to restore 4.5 million hectares by 2030, creating an urgent need for reliable monitoring frameworks. Global 10-m-resolution land-use and land-cover (LULC) datasets derived from remote sensing are increasingly used to track progress in forest restoration. However, their reliability in heterogeneous, smallholder-dominated landscapes remains uncertain. This study evaluates the accuracy of three widely used global products, Esri Land Cover, ESA World Cover, and Dynamic World, against > 5000 points from Malawi's National Forest Inventory (NFI) data. Results show that Esri achieved the highest overall accuracy (65.6%, 95% CI 64.2–67.0), followed by ESA World Cover (61.9%, CI 60.5–63.3), while Dynamic World performed considerably lower (46.9%, CI 45.5–48.3). Class-specific performance revealed that tree cover was consistently well detected (user's accuracy > 88%), whereas cropland and built-up areas were systematically underrepresented, and Rangeland was overpredicted. As global calls for ecosystem integrity monitoring grow stronger, our findings show that global datasets, although valuable for broad and temporal coverage, cannot independently support national-scale FLR monitoring in mosaic landscapes such as Malawi. We recommend integrated monitoring frameworks that combine the global consistency of Earth observation with the credibility of national inventories and the contextual relevance of community-based data. Such approaches are essential to capturing the complexities of land-use mosaics and supporting more accurate, legitimate, and policy-relevant restoration planning.