Abstract

Surface Bidirectional Reflectance Distribution Function (BRDF) correction of spectral data (Li et al., 2010) has important applications to time series based analysis and classification. However, it has been reasonably proposed that the BRDF information itself can be used directly in the time series applications for land cover mapping and climate change etc. To use such data it is important to understand the characteristics of BRDF and its variation over different cover and climate conditions and how they relate to well-understood variations in spectral data in terms of the land cover characteristics and changes. Many studies have suggested that BRDF is related to the characteristics of land cover types (Brown de Colstoun and Walthall, 2006 and Jiao et al., 2011), especially to vegetation structure (height and cover) (Lovell and Graetz, 2002; Li et al., 2013) and also climate patterns. In this study, 10 years of MODIS BRDF data sets (MOD43A1) from 2002 to 2011 have been used to conduct an analysis using time series data for land cover data products available in Australia. The data have been averaged over individual years to remove the seasonal patterns and variation for reasons which were outlined in Li et al. (2013) and are briefly discussed later in this paper. Using the root mean square (RMS, the distance of the shape function from Lambertian which is a measure of its asymmetry) as a BRDF shape indicator, with the inter-annual data series the study has found that: • The average RMS for three bands (red, near-infrared and shortwave infrared) for each year is well correlated with Normalized Difference Vegetation Index (NDVI) if it is separated by land cover classes. Correlation coefficients R 2 range between 0.5-0.7. The RMS also varies significantly between land cover classes. • Inter-annual variation of RMS is small for typical vegetation classes, especially for classes with high vegetation cover. • If Normalized Difference RMS is used (called NDRMS, calculated using red and near-infrared bands and the same formula as NDVI), its relationship with NDVI is much stronger than that of RMS. Correlation coefficients R 2 are close to 0.9 for most of the years. Each land cover class has well defined NDRMS patterns. The separation is clearer than for the RMS patterns. • NDRMS seems quite sensitive to climate change as indicated by NDVI but the relationship over the 10 years in some classes is different from the overall relationship between classes averaged over all years. In vegetated classes, NDRMS has tended to increase in this way much more sensitively after the change from a long dry period to wet years, and most particularly after 2009. The sensitivity has apparently increased with class average NDVI. From the above, it has been concluded that: • Both RMS and NDRMS are able to differentiate land cover classes defined in the Australian Dynamic Land Cover Dataset (DLCD) series well. They both correlate well with spectral NDVI if the patterns are separated by land cover classes and averaged at least over individual years (removing intra-annual effects). • Both RMS and NDRMS can potentially be used as additional features to map land cover. However, NDRMS seems to be the more sensitive of the two. • However, confident and successful use of these features will need additional understanding of the sources of the variation and the information they bring compared with traditional spectral data. In particular, further studies are needed to understand the rising sensitivity in NDRMS compared with NDVI as cover and greenness increase and the previously reported (Li at al., 2013) questions concerning relative phases of intra-annual variation in NDRMS and RMS relative to NDVI.

Full Text
Published version (Free)

Talk to us

Join us for a 30 min session where you can share your feedback and ask us any queries you have

Schedule a call