Global mapping of fractional tree cover for forest cover change analysis

Fractional tree cover facilitates the characterization of forest cover changes using satellite data. However, there are still substantial challenges in generating fractional tree cover datasets that satisfy the requirements of interannual stability for forest cover change monitoring. In this study, a global annual fractional tree cover dataset, named as GLOBMAP Fractional Tree Cover, was generated from MODIS observations with a resolution of 250 m during the period of 2000-2022. The tree cover estimation was improved relative to conventional global tree cover mapping methods by developing highly discriminative input features and near global sampling training data. MODIS annual observations were realigned at the pixel level to eliminate phenological differences among regions. The realigned annual series were condensed into twelve features showing high separability between trees and herbaceous vegetation to reduce the dimension of features. A global covering training dataset, comprising 465.88 million sample points, was extracted across the globe through the aggregation and combination of forest/land cover maps from ESA WorldCover, GlobeLand30, PALSAR FNF, and ESRI Land Cover to improve the representativeness of training data. A feed-forward neural network was calibrated to predict fractional tree cover from MODIS data. The spatial pattern of the estimation results was generally consistent with the CGLS-LC100 product at global scale, with the average MAE in global vegetated areas of 12.50 %, while our dataset provided extended temporal coverage. The interannual stability of the estimated tree cover series was improved compared to MODIS vegetation continuous fields products for deciduous broadleaf forests, evergreen broadleaf forests, and mixed forests, with the global average value of mean absolute deviation (MAD) in tree cover series reduced by 5.8 %, 46.9 %, and 18.3 %, respectively. The estimation results were assessed using globally distributed validation data around BELMANIP 2.1 sites and those derived from the USGS circa 2010 global land cover reference dataset, leading to the R2 values of 0.93 and 0.73, MAE of 4.24 % and 10.55 %, and RMSE of 11.45 % and 17.98 %, respectively. This dataset could enhance the capability to monitor forest cover changes, particularly gradual changes such as forest recovery.