Wang, B; Jia, K; Wei, XQ; Xia, M; Yao, YJ; Zhang, XT; Liu, DY; Tao, GF (2020). Generating spatiotemporally consistent fractional vegetation cover at different scales using spatiotemporal fusion and multiresolution tree methods. ISPRS JOURNAL OF PHOTOGRAMMETRY AND REMOTE SENSING, 167, 214-229.

Fractional vegetation cover (FVC) is considered one of the most important vegetation parameters and is relevant to characterizing vegetation status and ecosystem function. An FVC with a fine spatial resolution of 30 m is essential for monitoring vegetation change and regional studies, while an FVC with a coarse spatial resolution of hundreds to thousands of metres plays an important role in global change studies. However, high spatial resolution data usually have low temporal resolution and are often affected by cloud cover. The objective of this study is to propose a practical way to generate spatiotemporally consistent FVC products at Landsat and Moderate Resolution Imaging Spectroradiometer (MODIS) scales, which are 30 m and 250 m, respectively. The geostatistical neighbourhood similar pixel interpolator (GNSPI) was first used to fill in the missing values caused by unscanned gaps and clouds/shadows on Landsat-7 Enhanced Thematic Mapper Plus (ETM + ) data and to generate spatially continuous Landsat reflectance. Then, the enhanced spatial and temporal adaptive reflectance fusion model (ESTARFM) was used to generate time series Landsat reflectance data with the same temporal resolution as that of Global LAnd Surface Satellite (GLASS) FVC generated from MODIS data. The high temporal resolution Landsat reflectance was preliminarily used to estimate FVC at the Landsat scale. Finally, MultiResolution Tree (MRT) was employed to fuse the Landsat FVC and GLASS FVC to generate spatiotemporally consistent FVC products at different scales. The results show that the missing Landsat-7 ETM + data were filled well and spatial texture features were well preserved. The temporal resolutions of the Landsat and GLASS FVC products became consistent with an interval of one day at most. After MRT fusion, most of the root mean square error (RMSE) between the GLASS FVC and aggregated Landsat FVC dramatically decreased. The accuracy of the Landsat FVC validated by the ground-measured FVC improved after MRT fusion (before MRT: RMSE = 0.1031, R-2 = 0.9172, bias = - 0.0697; after MRT: RMSE = 0.0958, R-2 = 0.9173, bias = - 0.054). In addition, in the GNSPI-filled unscanned gaps and the ESTARFM-generated images, the Landsat FVC accuracy also improved slightly (before MRT: RMSE = 0.1065, R-2 = 0.9011, bias = - 0.0644; after MRT: RMSE = 0.1022, R-2 = 0.9023, bias = - 0.051). The accuracy of the GLASS FVC also improved (before MRT: RMSE = 0.0913, R-2 = 0.884, bias = - 0.0504; after MRT: RMSE = 0.0673, R-2 = 0.9483, bias = - 0.0444). Therefore, MRT could decrease the inconsistencies of different scales and reduce uncertainties in the FVC. In addition, MRT could fill in the missing data of the Landsat FVC directly, but there were a certain number of outliers in the fusion results, and the spatial transition was poor.