A multi-metric assessment of publicly available earth observation datasets reveals discrepancies in forest cover estimates in Paraguay, Zambia and Zimbabwe: higher resolution is not always a good indicator for accuracy

Unofficial AI-generated summary based on the public title and abstract. Not an official translation.

🔬研究者:This paper provides a reproducible framework for cross-dataset validation, highlighting the need for region-specific assessments in forest monitoring.

🏢実務担当者:Practitioners can use the multi-metric approach to select appropriate EO data for carbon accounting and forest management.

🏛政策担当者:Policymakers should be aware of significant discrepancies in forest cover estimates when designing carbon credit schemes and monitoring protocols.

Accurate forest monitoring is vital for climate mitigation, carbon credit schemes, and ecosystem management, particularly in tropical and subtropical regions facing rapid deforestation. Global Forest Watch (GFW) is widely used, yet has high commission errors; 45% in Sub-Saharan Africa, 17% in Latin America and 4% omission errors. This risks misleading policies and carbon accounting. To address this, we present a reproducible Google Earth Engine-Python workflow. It compares GFW with ESA WorldCover, Dynamic World, and Global-4 class PALSAR-2 across Paraguay, Zambia, and Zimbabwe. To ensure consistency, all products were standardised to forest/non-forest maps for pixel-based accuracy evaluation. Validation used 4.77 m Planet-NICFI mosaics, which provide high overall and F1 accuracies (87%–90%) and frequent temporal coverage. Their ability to capture seasonal clearing, and rapid regrowth offered a stronger reference. Visual interpretation of 500 random points further enhanced reliability over automated classification. These 500 points were overlaid on the forest/non-forest maps of each dataset to assess their agreement with the Planet-NICFI benchmark. Through this comparison, we derived multi-metric assessment results covering; overall accuracy, kappa coefficient, F1 scores, RMSE, and AUC. Dynamic World achieved the best performance in Paraguay, while GFW and Global-4 class PALSAR-2 performed better in Zambia and Zimbabwe. Importantly, the finer 10 m resolution of Dynamic World did not guarantee higher accuracy, underlining the need for region-specific assessments. Forest area calculations exposed further inconsistencies. In Paraguay, other datasets differ from GFW by approximately 3%–35%. In Zambia, deviations reach up to −47%. Zimbabwe shows the greatest divergence, with other datasets reporting 23%–70% less forest area than GFW. Comparisons with FAO statistics revealed additional discrepancies of −84% to +38%. These findings demonstrate that no single dataset can be assumed reliable across regions. Our framework provides a transparent, transferable approach that helps practitioners and policymakers select the most appropriate EO data for forest monitoring, carbon accounting, and environmental decision-making.

• semanticscholar https://doi.org/10.3389/frsen.2026.1679383first seen 2026-06-29 08:56:34