Upscaling fluxes from tower to landscape: Overlaying flux footprints on high-resolution (IKONOS) images of vegetation cover

Abstract

In this paper, we describe the process of assessing tower footprint climatology, spatial variability of site vegetation density based on satellite image analysis, and sensor location bias in scaling up to 1 km × 1 km patch. Three flat sites with different vegetation cover and surface heterogeneity were selected from AmeriFlux tower sites: the oak/grass site and the annual grassland site in a savannah ecosystem in northern California and a slash pine forest site in Florida, USA. The site vegetation density was expressed in terms of normalized difference vegetation index (NDVI) and crown closure (CC) by analyzing the high-resolution IKONOS satellite image. At each site, the spatial structure of vegetation density was characterized using semivariogram and window size analyses. Footprint maps were produced by a simple model based on the analytical solution of the Eulerian advection–diffusion equation. The resulting horizontal arrays of footprint functions were then superimposed with those of NDVI and CC. Annual sensor location biases for the oak/grass and the pine forest sites were <4% for both NDVI and CC, requiring no flux corrections in scaling from tower to landscape of 1 km². Although the annual grassland site displayed much larger location biases (28% for NDVI, 94% for CC), their temporal changes associated with averaging time showed a real potential to develop algorithms aimed at upscaling tower fluxes to the landscape in an effort to provide validation data for MODIS products.

Introduction

Natural vegetation is spatially heterogeneous. In particular, the spatial variability of vegetation density influences the lower atmospheric circulation and surface exchange of energy, water and carbon over a wide range of scales (e.g., Shen and Leclerc, 1995, Buermann et al., 2001, Cosh and Brutsaert, 2003). The scaling process involves taking spatial, temporal and process information at one scale and using it to derive information at another scale (Jarvis, 1995). In this process, the extraction of ecophysiologically significant information depends on spatial and temporal scales at which data are collected (Rahman et al., 2003). For example, field researchers employing micrometeorological flux measurement techniques are constrained by the predetermined spatial (<10⁴ to 10⁶ m²) and temporal (e.g., hours to years) resolution of the tower flux footprint (Baldocchi et al., 2001, Falge et al., 2002, Foken and Leclerc, 2004). On the other hand, multi-temporal, coarser remote sensing data allow mapping of the approximated surface flux over a wider region (>10⁶ to 10¹² m²) and longer time period (from days to decades). At present, bridging the gap between these two scales is a major challenge facing the research community. Accurate assessments of regional- and global-scale changes in the biosphere depend on the definition of practical scaling logic relevant to current flux sites, logic which incorporates a combination of field measurements, remote sensing, and ecological modeling.

The footprint of turbulent flux measurements characterizes its spatial structure that varies with wind direction, surface roughness, measurement height, and atmospheric stability (e.g., Leclerc and Thurtell, 1990; Schmid, 2002). Recent development of footprint models provides diagnostic tools to quantify the representativeness of tower flux measurements for selected sites (e.g., Schuepp et al., 1990, Leclerc and Thurtell, 1990, Horst and Weil, 1994, Schmid, 1997, Baldocchi, 1997, Amiro, 1998, Schmid and Lloyd, 1999, Leclerc et al., 2003a, Leclerc et al., 2003b, Soegaard et al., 2003, Foken and Leclerc, 2004, Levy et al., 2004). Despite many current studies on detailed footprint modeling and experimental validation (e.g., Leclerc and Thurtell, 1990, Wilson and Swaters, 1991, Horst and Weil, 1994, Finn et al., 1996, Leclerc et al., 1997, Leclerc et al., 2003a, Leclerc et al., 2003b; Cooper et al., 2003), the temporal and spatial variability of footprints has not yet been investigated and the associated influence of varying site vegetation density on tower flux measurements. One of the practical problems in using a footprint model as an operational tool is that the source contribution in the area of a prospective measurement site is not known a priori. Recently, it has been demonstrated that long-term patterns of source contributions (i.e. ‘footprint climatology’) provide essential information about the vegetation sampled when measuring long-term fluxes especially over heterogeneous landscapes (e.g., Amiro, 1998, Schmid and Lloyd, 1999, Stoughton et al., 2000, Levy et al., 2004). The footprint climatology can be combined with information on the spatial variability of vegetation density characteristics provided by satellite image analysis. The current remote sensing technology provides high-resolution images of vegetation density in the form of the normalized difference vegetation index (NDVI) and crown closure (CC). The use of both the very high-resolution IKONOS imagery and the in situ flux footprint should result in more accurate validation data for process models and MODIS products.

Our objective is to examine the representativeness of tower fluxes in scaling up to the scale of satellite images (1 km × 1 km) by overlaying information of spatial variability of vegetation density on that of the flux footprint. The null hypothesis tested is that there is no significant difference in the flux indices used (NDVI, CC) between those averaged over the satellite domain and those selected and weighted by the footprint criteria. Accordingly, we selected three AmeriFlux sites with different vegetation densities to examine the influence of patch-scale heterogeneities on flux footprints of eddy-covariance towers. High-resolution IKONOS satellite images are used to determine NDVI and CC in the vicinity of the three sites. Footprint maps were produced using an analytical solution of the two-dimensional Eulerian advection–diffusion equation (Horst and Weil, 1994). The resulting horizontal arrays of footprint functions are then superimposed with those of NDVI and CC. The result is a ‘tower location bias’ (Schmid, 1997) which should be taken into account with the use of remote sensing and biosphere models to scale tower fluxes and field measurements. The present study also examines the spatial structure of vegetation density at the sites using semivariogram and window size analyses. Finally, 16 days averages of sensor location biases for NDVI and CC were estimated throughout the year and their temporal changes related to both the MODIS data product and the gross primary production from eddy-covariance flux measurements.