Abstract

Abstract Fractional tree cover facilitates the depiction of forest density and its changes. However, it remains challenging to estimate tree cover from satellite data, leading to substantial uncertainties in forest cover changes analysis. This paper generated a global annual fractional tree cover dataset from 2000 to 2021 with 250 m resolution (GLOBMAP FTC). MODIS annual observations were realigned at the pixel level to a common phenology and used to extract twelve features that can differentiate between trees and herbaceous vegetation, which greatly reduced feature dimensionality. A massive training data, consisting of 465.88 million sample points from four high-resolution global forest cover products, was collected to train a feedforward neural network model to predict tree cover. Compared with the validation datasets derived from the USGS circa 2010 global land cover reference dataset, the R 2 value, MAE, and RMSE were 0.73, 10.55%, and 17.98%, respectively. This dataset can be applied for assessment of forest cover changes, including both abrupt forest loss and gradual forest gain.