Estimation and validation of 30 m fractional vegetation cover over China through integrated use of Landsat 8 and Gaofen 2 data

Abstract

Fractional vegetation cover (FVC) is an important parameter that reflects the status of vegetation and can be used to quantify vegetation dynamics in system models of Earth. Currently, global FVC products are usually available at coarse spatial resolutions, and so cannot satisfy the requirements for detailed investigations of the distribution of vegetation cover in different regions worldwide. Thus, a fine-resolution FVC product that covers a large spatial extent is needed. In this study, we develop an operational framework to produce Landsat-like spatial resolution FVC products using high-resolution images and a machine learning algorithm. We used 1 m Gaofen 2 data to calculate the 30 m FVC, and then applied as training and testing data for modeling. The random forest regression model using Landsat 8 surface reflectance and solar angle as inputs outperformed the other models in terms of accuracy and efficiency and was selected to estimate the 30 m FVC. The model has a high training accuracy, with an R2 of 0.978 and an RMSE of 0.042. When validated using 346 independent ground measurements from 14 sites around the globe, the R2 was 0.814 and the RMSE was 0.170. We compared three coarse-resolution FVC products at both regional and national scales, including GLASS, GEOV2 and CGLOPS FVC products. The 30 m Landsat FVC from this study was spatially and temporally consistent with all reference FVC products, but provided advantages in revealing the fine spatial details of vegetation cover. The performance of the 30 m FVC model varied with vegetation type, but showed the highest consistency with the GLASS FVC for most vegetation types. Finally, by applying the established 30 m FVC model to all Landsat 8 data over China from 2013 to 2018, we derived a seasonal maximum FVC mosaic, which successfully demonstrated the seasonal and spatial variations in vegetation cover in China. We conclude that the proposed framework enables accurate estimation of FVC at a high resolution over a large spatial scale and so, can be used as part of an operational approach for generating a global fine-resolution FVC product.