A comparison of methods for estimating fractional vegetation cover in arid regions

We compared a set of methods for estimating the fractional vegetation cover (f(c)) of sparse desert vegetation over an arid region of southern Xinjiang, China. Six kinds of remote sensing inversion models (an NDVI regression, a spectral mixture analysis (SMA), a pixel dichotomy model, a three-band maximal gradient difference (TGDVI) model and two modified TGDVI models) were used to derive fc from remote sensing data, and the results were compared with f(c) values measured in the field to select an appropriate model to derive the fractional cover of sparse desert vegetation in arid regions. The NDVI regression based on field f(c) and the NDVI for the sampled pixels in September 2006 showed the highest precision, while the results of 2007 showed that the NDVI regression method is inappropriate for depicting vegetation characteristics in other growing season because the empirical model highly depend on the specified in situ measurement. The SMA approaches yielded higher precision than the other models, indicating that it is applicable for analysing the coverage of sparse desert vegetation. The pixel dichotomy model can yield a high precision based on finely detailed vegetation maps. However, it requires the measurement of many parameters. The TGDVI model is simple and easy to implement, and the values that it predicted for the coverage of high-density vegetation and barren areas were close to those measured in the field, but theft values of sparsely vegetated areas were underestimated. The predictions of the modified TGDVI models were close to the values measured in the field, indicating that these modified models can reliably and effectively extract information on the fractional cover of sparse vegetation in an arid region. We analyzed the models' sensitivity with respect to rainfall because the short-wavelength infrared bands used in the two TGDVI models proposed in this study are sensitive to moisture. The results showed that the modified TGDVI models' accuracy was not affected by increasing soil moisture content caused by rain. However, the NDVI regression, SMA and TGDVI were sensitive to the change of soil moisture content. Moreover, the two modified TGDVI models yielded negative values for water sources, such as reservoirs and rivers, implying that they are effective for characterising water bodies. However, the modified TGDVI models cannot predict f(c) in snow- and glacier-covered regions, producing abnormally high rather than zero values. Additionally, the predictions before and after snowfall on the top of a mountain show a linear increasing relationship, suggesting that the short-wavelength infrared band may be useful to predict snow depth. (C) 2011 Elsevier B.V. All rights reserved.