Evaluation of spectral unmixing techniques using MODIS in a structurally complex savanna environment for retrieval of green vegetation, nonphotosynthetic vegetation, and soil fractional cover

Abstract

This study tests the performance of spectral mixture analysis (SMA) and multiple endmember spectral mixture analysis (MESMA) for estimation of green vegetation (GV), nonphotosynthetic vegetation (NPV), and soil fractions in the heterogeneous, structurally complex savannas of the western Kalahari using the Moderate Resolution Imaging Spectroradiometer (MODIS) nadir-bidirectional reflectance distribution function (BRDF) adjusted reflectance (NBAR) product. Extensive fieldwork took place during the dry and wet seasons of 2009 to 2011 at 15 sites distributed along a 950km long transect, stretching across western Botswana, each site was visited once during the wet season and once during the dry season. Data collection included a traditional line-intercept transect (LPIT) and a new spectral line point intercept transect method (SLPIT) were used to test the performance of a variety of unmixing procedures (MESMA vs. SMA) and endmember models. The results for this structurally complex landscape are consistent with results from similar studies undertaken in more homogeneous areas. GV cover was retrieved much more accurately than NPV or soil cover. MESMA also produced estimates of fractional cover with less error than simple SMA. However, the errors observed are greater than those observed for more homogeneous environments. Unlike the line-point intercept method, which requires user interpretation of vegetation greenness, the new method uses spectral data collected across the entire reflected solar spectrum to derive estimates of GV, NPV and soil fractional cover through spectral unmixing. Our results show that that the SLPIT fractions generally agree better with remotely-derived fractions than the LPIT-derived fractions. However, remote sensing of GV, NPV, and soil fractional cover, especially in heterogeneous landscapes and at the spatial resolution of MODIS, remains challenging. Nonetheless, the data do show that at this resolution various unmixing methods have the potential to inform our understanding of ecosystem dynamics in these environments.