Generating spatiotemporally consistent fractional vegetation cover at different scales using spatiotemporal fusion and multiresolution tree methods

Abstract

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, R² = 0.9172, bias = −0.0697; after MRT: RMSE = 0.0958, R² = 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, R² = 0.9011, bias = −0.0644; after MRT: RMSE = 0.1022, R² = 0.9023, bias = −0.051). The accuracy of the GLASS FVC also improved (before MRT: RMSE = 0.0913, R² = 0.884, bias = −0.0504; after MRT: RMSE = 0.0673, R² = 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.

Introduction

Fractional vegetation cover (FVC) is defined as the proportion of the vertical projected area of green vegetation to the total statistical area, which is recognized as one of the most important parameters for describing terrestrial ecosystems and monitoring vegetation conditions (Gao et al., 2020, Wang et al., 2018). FVC has a significant influence on the exchange of carbon, water and energy at the land surface (Wang et al., 2017b). FVC has also been applied to many land surface process simulations and global change studies, such as soil erosion monitoring, drought monitoring, desertification evaluation, hydrological simulation, agricultural monitoring (de Asis and Omasa, 2007, Jia et al., 2016, Matsui et al., 2005, Zhang et al., 2013), the Earth’s energy balance and climate change (Gutman and Ignatov, 1998, Roujean and Lacaze, 2002). Therefore, accurate, timely and continuous FVC products at global and regional scales are of great significance.

Remote sensing is the most effective way to estimate FVC due to its advantages of extensive coverage and repeated observations (Liang et al., 2012). However, there is a contradiction between the temporal resolution and spatial resolution of remote sensing data; data with fine spatial resolution usually have a long revisit cycle, and vice versa (Bolton and Friedl, 2013). Currently, there are several large-scale FVC products with coarse spatial resolution, such as the Envisat Medium Resolution Imaging Spectrometer (MERIS) (Bacour et al., 2006), the Carbon Cycle and Change in Land Observational Products from an Ensemble of Satellites (CYCLOPES) (Baret et al., 2007), the Geoland-2/BioPar version 1 (GEOV1) (Baret et al., 2013) and the GLASS (Jia et al., 2015) FVC products. The spatial resolutions of these products range from a few hundred metres to several kilometres. High temporal resolutions within 10 days allow these FVC products to capture rapid changes in vegetation on the land surface. Furthermore, the spatial-temporal continuity of these products makes them capable of monitoring vegetation change over large areas and conducting global change analysis (Yang et al., 2018, Yu et al., 2018). However, coarse spatial resolution data are often inadequate for highly heterogeneous areas such as agricultural landscapes because they cannot capture the detailed spatial distribution and vegetation variation patterns, which are crucial for crop classification, crop growth monitoring and yield estimation (Azzari and Lobell, 2017, Doraiswamy et al., 2004). Therefore, high spatial resolution data are vital for regional studies, such as precision agriculture and ecosystem function monitoring (Lobell and Asner, 2003, Röder et al., 2008).

Landsat data have a fine spatial resolution of 30 m, providing sufficient details for land surface variation, which is an appropriate scale to reflect human activities (Gao et al., 2015). However, the 16-day revisit cycle and frequent cloud contamination make it difficult to directly utilize Landsat data to detect rapid vegetation change, such as monitoring the critical periods of crop growth (Bolton and Friedl, 2013). In fact, there are few cloud-free data in most areas, and some of them have to be discarded due to heavy cloud cover, which causes the discontinuity of land surface information in time series and space. Moreover, sensor failures (such as ETM+) make the application of Landsat data more challenging (Zhu et al., 2012a). Therefore, FVC estimated from high spatial resolution data is usually spatially and temporally discontinuous, which limits FVC application in land surface process simulations and ecosystem modelling research (Bian et al., 2017).

Many spatiotemporal fusion algorithms have been developed to weaken the influence of clouds and obtain high spatial and high temporal observations with Landsat satellites. The spatial and temporal adaptive reflectance fusion model (STARFM) takes advantage of the high temporal resolution of MODIS and high spatial resolution of ETM+ to produce high spatial resolution and high frequency data (Feng et al., 2006). Zhu et al. (2010) improved the STARFM and proposed the ESTARFM algorithm by using two pairs of high- and low-resolution images, which produces a more accurate prediction of surface reflectance in heterogeneous landscapes and has been widely used (Dong et al., 2016, Fu et al., 2014, Yan et al., 2018). In regard to the unscanned gaps of Landsat-7 ETM+, several algorithms have been proposed to fill the missing data, such as the neighbourhood similar pixel interpolator (NSPI) (Chen et al., 2011) and GNSPI (Zhu et al., 2012b), which can recover the missing spectral information with good precision. Although such spatiotemporal algorithms can increase the spatial and temporal continuity while observing fine land surface variations compared with that of medium spatial resolution data, uncertainties will also be introduced to the reconstructed surface reflectance, which are caused by a series of assumptions that simplify the complex land surface conditions. GNSPI assumes that neighbouring similar pixels have similar temporal changing patterns (Zhu et al., 2012b), while ESTARFM assumes that reflectance changes linearly during a short period and that both the proportion and reflectance change rate of each endmember are stable (Zhu et al., 2010). Therefore, the accuracy of subsequent FVC estimation will be affected (Deng et al., 2019). In addition, different FVC estimation methods commonly lead to different results due to different algorithm mechanisms, such as empirical methods and pixel unmixing model and physical-based methods (Jia et al., 2017, Jiapaer et al., 2011, Zou et al., 2018). Moreover, FVC estimates derived from multiresource satellite data are usually spatially and temporally inconsistent, particularly across different spatial resolutions. Therefore, generating spatiotemporally consistent FVC products across different scales is of great significance and potential for related research.

To reduce the uncertainties and overcome the problem of data inconsistency across different spatial resolutions, an MRT method has been developed based on the assumption that a statistical model is autoregressive in its levels of resolution (Chou et al., 1994b). MRT has been used for a series of remote sensing issues, especially for estimating satellite-based variables using mass data because it can provide an optimal estimation with efficient computation. Parada and Liang (2004) assimilated different resolution near-surface soil moistures and demonstrated that MRT could recover relevant spatial features and significantly reduce the RMSE. He et al. (2014) fused three surface albedo products from Multiangle Imaging Spectroradiometer (MISR), MODIS and Landsat to generate consistent albedo data at different spatial resolutions. Shi et al. (2016) integrated advanced spaceborne thermal emission and reflection radiometer (ASTER) and GLASS broadband emissivity (BBE) products to obtain better BBE products. However, few studies have explored the potential of MRT for integrating high and medium spatial resolution vegetation parameters, including FVC. Considering the advantages of MRT in improving accuracy, reducing uncertainty and minimizing bias across different spatial resolutions, this method is explored to generate spatiotemporally consistent FVC products from Landsat FVC and GLASS FVC products in this study.

There are two objectives in this study: (1) reduce the spatial and temporal inconsistency between Landsat and GLASS FVC products and generate spatiotemporally consistent FVC products at 30 m and 250 m scales; and (2) improve the accuracy of FVC products, especially for the data reconstructed by the spatiotemporal fusion algorithm. To achieve these goals, the unscanned gaps of ETM+ were first filled using the GNSPI method, and then the ESTARFM was used to generate temporally consistent reflectance data between Landsat and MODIS, which was subsequently used to estimate FVC. Finally, MRT was investigated to produce spatially consistent FVC products across different spatial resolutions. In contrast, the missing Landsat FVC data were merged by GLASS FVC directly without filling in the unscanned gaps of Landsat reflectance data to test the ability of MRT to directly fill in missing values.