Multi-Metric Accuracy Assessment of Forest Cover Datasets in Three Countries

500 validation points were used to assess the accuracy of four major earth observation forest cover datasets in Paraguay, Zambia, and Zimbabwe. The analysis reveals significant discrepancies, with forest area estimates from other datasets deviating from Global Forest Watch by up to 70% in Zimbabwe. This assessment was conducted by Natasha Gapare and published in 2026.

Use Cases

• Compare overall accuracy, kappa coefficient, F1 scores, RMSE, and AUC metrics between GFW, ESA WorldCover, Dynamic World, and Global-4 class PALSAR-2.
• Analyze region-specific performance, such as Dynamic World's best performance in Paraguay versus GFW's better results in Zambia and Zimbabwe.
• Assess commission and omission error rates, like the 45% commission error for GFW in Sub-Saharan Africa, for specific policy applications.
• Calculate and reconcile forest area estimates, which showed deviations from GFW ranging from -47% in Zambia to -70% in Zimbabwe.
• Validate earth observation products against a 4.77m Planet-NICFI mosaic reference, which achieved 87%–90% overall and F1 accuracies.

Strengths

• Validation uses 500 manually interpreted random points for reliability over automated classification.
• Benchmark Planet-NICFI mosaics provide high-resolution (4.77m) reference data with 87%–90% overall and F1 accuracies.
• Multi-metric assessment includes overall accuracy, kappa, F1, RMSE, and AUC for a thorough comparison.

Limitations

• The dataset's scope is limited to only three countries: Paraguay, Zambia, and Zimbabwe.
• The core validation set is relatively small, consisting of 500 sample points across all assessed regions.
• The underlying data file is a 29.8 KB DOCX document, indicating a summary report rather than a primary data table.

Provenance

Source

Analysis by author Natasha Gapare, comparing datasets from Global Forest Watch, ESA WorldCover, Dynamic World, and JAXA's Global-4 class PALSAR-2.

Collection Method

A reproducible Google Earth Engine-Python workflow standardized products to forest/non-forest maps for pixel-based accuracy evaluation against Planet-NICFI mosaics.

Freshness

Data was last updated in April 2026.

Geography

Paraguay, Zambia, and Zimbabwe.