Accuracies, discrepancies, and challenges of the 10 m global land cover products in mountains
Amin Naboureh, Ainong Li, Jinhu Bian, Meysam Moharrami, Hamid Ebrahimy, Guangbin Lei, Xi Nan, Zhengjian Zhang, Bakhtiar Feizizadeh, Paolo Dabove, Mohsen Makki, Tobias Sauter, Sara Attarchi, Meisam Amani
Abstract
The integration of advanced remote sensing technology, machine learning, and cloud computing platforms has resulted in a new era of high-resolution (10 meter) Global Land Cover (GLC) products. These GLC products often adopt varying methodological frameworks, leading to potentially inconsistent classification outcomes that hinder their usability. This challenge could particularly be more pronounced in mountainous areas because their complex terrain and frequent mountain/cloud shadows make accurate Land Cover (LC) classification difficult. To investigate this issue, we assessed the reliability of three 10-m GLC products: ESRI Land Use/Land Cover, ESA WorldCover, and Dynamic World (DW). We examined their thematic accuracy and spatial and area consistency at the global, continental, and major mountain range levels (24 ranges in total). Our statistical assessment showed that ESRI demonstrated the highest overall accuracy (68.5 ± 0.42%) at the global mountain level, with a small lead over ESA (66.1 ± 0.42%) and DW (65.6 ± 0.45%). However, none of the three GLC products consistently excelled at continental and major mountain levels, underscoring significant regional variability. While water class exhibited F1-scores of more than 85%, other classes (e.g. mixed vegetation and cropland) showed F1-scores of between 40% and 70%. Additionally, we noted substantial discrepancies in the area estimates for key LC classes. Notably, only 53% of pixels had perfect agreement among products at the global scale, with particularly poor consistency across complex terrains such as the Tian Shan and Zagros mountains. Overall, this study provides actionable insights for users of GLC data in global mountainous regions and emphasizes the urgency of addressing classification inconsistencies in future products. To support future research and validation efforts, we provide open access to the dataset of approximately 56,000 validation sample points collected in this study at https://zenodo.org/records/15354081.