Fractional vegetation cover estimation in southern African rangelands using spectral mixture analysis and Google Earth Engine

Highlights

• Grasslands require continuous and dynamic condition data for sustainable management.

• Fractional cover as indication of grass cover (%) for degradation monitoring.

• Spectral unmixing for fractional cover estimation is accurate and transferable.

• Endmember class subdivision maximises separability.

• SWIR2, NDVI, EVI, MSAVI2 and DBSI is the optimal feature combination for Sentinel-2.

Abstract

Grasslands are under continuous threat of conversion and subsequent degradation, which has a detrimental impact on grassland productivity and grazing capacity, affecting the livestock industry. Fractional vegetation cover as indicator of grassland condition and productivity has been extensively researched, however, existing approaches and products are limited with respect to accessibility, affordability, applicability, and transferability. This study evaluated the use of publicly available satellite imagery, spectral mixture analysis and cloud geoprocessing technologies for dynamic, continuous, and accurate estimation of FVC for sustainable management. A linear spectral mixture model was developed, calibrated, and implemented in Google Earth Engine using Sentinel-2 and Landsat 8 imagery. Model accuracy and spatial and temporal transferability were evaluated using existing benchmark products and field data. It was found that Sentinel-2 performed the best using a feature combination of the SWIR2 band and the NDVI, EVI, MSAVI2 and DBSI indices. Accuracies were further improved by dividing the woody and bare endmembers into subclasses. The approach proved both spatially and temporally transferable, thus this research provides a robust approach to FVC estimation using limited field data and open source remote sensing imagery. The combination of this research with further grassland productivity modelling could prove valuable for sustainable environmental and economical rangeland planning and management.

Introduction

Grasslands are one of the most important land uses in South Africa and are not only critical ecosystems service providers, but also play a vital role in water production and agriculture (Driver, et al., 2011). Despite their importance, between 60 and 80% of South Africa’s grasslands have been irreversibly transformed (Acocks, 1953), and the biome has been identified as critically endangered (Wang et al., 2017). This grassland transformation can lead to irreversible land degradation, which consequently has a detrimental effect on grassland productivity. Such a decline in grassland productivity poses challenges for the livestock industry, a major food provider of the South African agricultural sector, which relies heavily on managed and indigenous grasslands for grazing of sheep, goats, and cattle. One of the main drivers of large-scale grassland conversion and degradation is woody plant encroachment (WPE) (Shroder et al., 2016), defined as the proliferation of woody plants in grasslands, savannas and rangelands (Archer et al., 2017). The occurrence of WPE is often associated with livestock overgrazing and browsing and the two interrelating factors should both be considered when investigating the productivity of savannas and rangelands.

To ensure sustainable economic and environmental management of grasslands, the livestock industry and producers require regular, high quality land condition data to allow the identification, analysis, and potential prevention of degradation occurrences and to provide quantification of the detrimental impacts on grassland productivity. Current research on the estimation of grassland and rangeland condition and productivity using remote sensing (RS) typically make use of various biophysical indicators (Jiménez-Muñoz et al., 2009), such as normalised difference (NDVI), leaf area index (LAI), fraction of absorbed photosynthetic radiation (fAPAR), evapotranspiration (ET) and aboveground standing biomass (ASB). Fractional vegetation cover (FVC) is one of the main, but often overlooked, biophysical parameters relating to landscape surface processes (Guerschman et al., 2009) and involves determining the proportional area of specific types of groundcover a pixel consists of, based on various spectral characteristics of the pixel (Jiménez-Muñoz et al., Feb. 2009). The types of groundcover differ depending on the FVC application, but usually include green (e.g. leaves), dead (e.g. wood) and bare (e.g. soil). This essentially allows discrimination between productive (grass) and unproductive (woody branches, bare soil) groundcover within a pixel, which in turn is potentially useful for deriving subsequent grazing and browsing productivity estimates.

Estimating an accurate FVC is a complex process, as it involves extracting sub-pixel information from temporally, spatially and spectrally variable phenomena such as vegetation (Li et al., 2014). Jimenez Munoz (Jiménez-Muñoz et al., 2009) highlighted the three most frequent techniques in literature for retrieving FVC: vegetation indices (VIs), machine learning and spectral mixture analysis (SMA). The VI approach involves collecting field measurements of FVC components, e.g. bare soil and vegetation, identifying corresponding VI values and deriving regression relationships (Gitelson et al., 2002). Although this approach can provide accurate estimations, the established relationships are site-specific and can often not be transferred to other regions without recalibration of parameters (Liang et al., 2012). As computer technology improves, the machine learning approach for FVC estimation has become an increasingly popular alternative (Yang et al., Aug. 2016, Shiferaw et al., Mar. 2019, Jia, Sep. 2015). Common machine learning algorithms used for FVC include support vector machines, decision trees and neural networks. Successful application, however, requires a large number of training points (Campbell, 2002), which is often time-consuming and expensive to obtain.

The SMA approach, which is the focus of this study, involves extraction of land cover (LC) information from satellite imagery at a sub-pixel level. Endmembers, which represent the pure spectral characteristics of a specific LC feature (Plaza et al., 2004), are used to divide the pixel into its different components. Various models exist for performing SMA, of which the linear spectral mixture model (LSMM) is the simplest and most common. Relatively few endmembers are required to describe the surface composition per pixel, thus an LSMM is an appropriate choice when field data is limited. When performing SMA, the main problem is identifying and extracting high quality, pure endmember spectra that accurately reflect the spectral profile of each endmember. In a study comparing several methods of endmember extraction in deriving FVC for an agricultural region in Spain (Jiménez-Muñoz et al., 2009), the manual method of using a land use map produced the best results.

A number of public global-scale FVC products have been created, most well-known the MODIS Vegetation Continuous Fields (VCF) product (MOD44B collection 6) with a 250 m spatial resolution and a yearly temporal resolution (Townshend et al., 2017). Another popular product is the Copernicus Dynamic Land Cover (CDLC) product (Tsenbazar et al., 2018) created on a once-off basis for 2015 at 100 m. Although both products made use of a machine learning approach, the two products differ with respect to spatial resolution and the groundcover types and definitions used. CDLC provide fractional cover layers for 10 groundcover types: shrubland, forest, moss and lichen, snow, bare/sparse vegetation, permanent water, seasonal water, cropland, built-up and herbaceous vegetation. MOD44B, however, provides only three i.e. tree, non-tree and non-vegetated (bare), while shrubs are included in the non-tree class.

In southern Africa, both vegetation indices (Scanlon et al., 2002) and spectral unmixing (Gessner et al., 2013), as well as a combination of these approaches (Sankaran, 2005), have been used to estimate FVC. A woody fractional canopy cover for South Africa was also developed by the Council for Scientific and Industrial Research (CSIR) Ecosystems Earth Observation Unit as part of the Carbon Sinks Atlas project (Department of Environmental Affairs, 2017). The 100 m and 1 km savanna woody fractional cover was created for the year 2011 with the use of Synthetic Aperture Radar (SAR) and Light Detection and Ranging (LiDAR) (Naidoo et al., 2016). More extensive studies have been performed in Australia, where rangelands and grasslands have very similar vegetational structure and climatic conditions to that of South Africa. Of particular interest is the FVC product developed by AusCover based on the methodology of Scarth et al. (2010). Scarth et al. (2010) developed a non-negative constrained least squares linear spectral unmixing model to derive an accurate FVC from Landsat 8 imagery for the extent of Australia. In contrast to the previously mentioned studies, this study makes use of medium resolution satellite imagery (30 m spatial resolution) to produce a fractional cover layer, thus eliminating the need for expensive hyperspectral or very high resolution data to produce accurate FVC results. The methodology has recently been adapted to determine FVC using Sentinel-2 imagery as well, which combines the advantages of both high temporal and high spatial resolution data for continuous monitoring of rapid changes in degrading grassland ecosystems (Zhang et al., Sep. 2017). A drawback of this product for grazing management is the FVC class definitions, namely bare soil (BS), photosynthetic vegetation (PV) and non-photosynthetic vegetation (NPV) (Scarth et al., 2010). The PV class contains both tree and shrub leaves, as well as herbaceous vegetation, thus combining non-grazeable and grazeable material in the same class.

Many of these studies required either hyperspectral data, very high resolution imagery or very expensive active sensor technology (SAR, LiDAR) to produce accurate FVC estimates, which introduces challenges with regards to accessibility, availability and affordability. Those that did not, typically provide FVC estimates at global scale, which is too coarse for identifying complex vegetational patterns and dynamics in savannas and woody encroached grasslands. In addition, the groundcover types and definitions used as fractional cover classes differ widely, complicating the process of comparing different FVC products. This need for high spatial and temporal resolution FVC using publicly accessible RS data that is suitable for local grazing potential analysis was identified as a gap in literature. This study aims to address this gap by developing a dynamic, continuous and accurate estimation of FVC for grazing purposes through the combination of publicly available satellite imagery, spectral mixture analysis and cloud technologies. An LSMM is developed, calibrated and implemented in Google Earth Engine, a geoprocessing cloud platform that provides continuous, dynamic processing functionality at various scales (Gorelick et al., Jul. 2017). The FVCs produced using the LSMM are evaluated using existing benchmark products and field data. The spatial and temporal transferability of the LSMM is also assessed to determine whether it can form part of a robust grassland and rangeland management system. The results are interpreted within the context of finding an appropriate approach to FVC estimation for operational grassland and rangeland degradation management for the livestock industry.