GHLC: global harmonized land cover datasets for 2020–2024 at 10 m and 30 m resolution based on a rule-based ensemble approach

Abstract. The paper describes the production and assessment of GHLC: a global harmonized land cover product at 10 meter spatial resolution for 2020 and at 30 meter resolution for 2020-2024. GHLC is generated using a rule-based ensemble approach that integrates 24 individual land cover and auxiliary layers from different projects. Because each class label is assigned based on an explicit, documented rule rather than a trained model, every classification decision in GHLC can be traced to a specific input dataset and threshold, making the product transparent, reproducible, and straightforward to update as improved input datasets become available. These 24 layers are combined around a primary spatial backbone provided by ESA WorldCover and GLAD GLCLUC (Global Land Cover and Land Use Change). The documented sequence of conditional rules describes how other thematic datasets (e.g., Global Mangrove Watch for mangroves, WorldCereal for maize, South American maps for soybeans, and GLC_FCS30D for detailed forest subtypes) are fused with ESA WorldCover and GLAD GLCLUC. The target legend is structured hierarchically to balance thematic detail with global mapping feasibility (Level 1 consists of 10 broad classes; Level 2 consists of 40+ distinct classes, including specific crop types, varying grassland management types, and forest types). The map was validated against a large pool of independent reference samples from the Global Ensemble Land Cover (GELC) database, matched to each product's mapped years. At the broadest level, the GHLC product achieved an overall accuracy of 0.92 for the 10 m and 0.86 for the 30 m maps, respectively. Forests, forest plantations, and woodland (F1 ≥ 0.94) and water bodies (F1 ≥ 0.85) performed best across both resolutions due to their high spectral distinctiveness and spatial coherence. Cropland classification was more accurate in the 30 m product (F1 = 0.81) than in the 10 m product (F1 = 0.72), as the 30~m GLAD GLCLUC backbone provides a stronger cropland signal. Macro-averaged F1 across all classes was effectively identical between 10 m (0.63) and 30 m (0.62), demonstrating that spatial resolution alone does not govern overall accuracy. The poorest performance was registered for sparsely vegetated ecosystems and inland wetlands. Urban ecosystems appear to be overestimated in the 30 m map. The datasets can be used for reporting on ecosystem extent and ecosystem services in ecosystem accounting. The biggest limitations of the GHLC are (1) that it is based on manual rules that may be subject to errors, and (2) that it relies on 24 independently derived external layers, each produced and maintained by separate research groups or agencies. If those external groups cease updating or maintaining their specific datasets, the corresponding GHLC rules cannot be applied, causing the classification system to degrade.