Estimation of High-Resolution Fractional Tree Cover Using Landsat Time-Series Observations

High-resolution fractional tree cover mapping allows for the representation of spatial details of tree distribution and contributes to forest ecosystem monitoring and modeling. However, the estimation of tree cover at high resolution is challenging, especially in sparsely tree-covered and mountainous areas interfered with background signals (e.g., soil, grass, and water) and terrain shadow. This article presents a fractional tree cover estimation algorithm from Landsat time-series observations at 30-m resolution, with the effects of background and terrain shadow reduced. The seasonal profiles of vegetation index and surface reflectance were constructed from multiyear Landsat data. Three phenological metrics were extracted from the seasonal profiles as input features for tree cover estimation. The training data were collected from the European Space Agency (ESA) WorldCover product and used to calibrate a feedforward neural network model to predict cover fractions. This algorithm extracted the tree cover of major forest types, including boreal, temperate, tropical dry, and moist forests. It also captured the sparse tree cover in areas containing mixtures of tree crowns and grass, bare soil, crop, impervious surface, and water. In two dense montane forest areas, the tree cover fractions on shady and sunny slopes were estimated consistently. The estimation results were evaluated through the reference samples generated from Google submeter-resolution image classification. The values of
$R$
-squared (
$R^{2}$
), root-mean-square error (RMSE), and mean absolute error (MAE) reached 0.78, 15.71%, and 11.09%, respectively. The proposed algorithm can be applied to monitor tree cover in spatially fragmented forest areas and sparse forests.