[1] However, data is not available with such frequency due to limitations on data transferring to Earth.

[2] Because the spectral signatures of dead or senesced vegetation and soil are many times quite similar over these wavelengths, these surface materials are hard to distinguish from one another using multispectral data alone, such as that acquired by Landsat and Spot sensors.

[3] Conversion of raw DN values to reflectance values consists of correcting for atmospheric scattering and incident solar radiation. This process will be explained further in the Methods section.

[4] Values below zero generally correspond to water or dark surface materials containing no vegetation at all. Highly vegetated areas generally have a value within a range of 0.4-1.0.

[5] Both nonlinear and linear mixture modeling have been used and studied. While nonlinear mixture modeling has been shown to accurately reflect physical arrangements of materials and has been used on a number of vegetation types, linear mixture modeling has been more widely used and has proved sufficiently accurate for extraction of vegetative land cover (Adams et al., 1993; Roberts et al., 1993; Smith et al., 1990 a&b.).

[6] It is important to note that the number of different land cover fractions that can be generated through SMA is limited by the number of bands in the data used. This is discussed in detail below.

[7] The same study found NDVI to indicate the correct sense of change only 67% of the time, showing SMA to be drastically superior (Elmore et al., 2000).

[8] Hyperspectral data is data acquired over more than 255 bands or wavelengths. This kind of data requires a large amount of processing capacity. However, it allows for a much more detailed analysis of surface features, as more spectral information is available with which to distinguish materials.

[9] It is important to note that this step can be extremely time consuming. However, it greatly improves the accuracy in the change detection measurements.

[10] Soil background effects were discussed earlier as being a limitation of NDVI based on earlier studies (Bausch, 1993; Huete, 1985; Huete, 1988; Huete et al., 1991;  Major et al., 1990; Oi et al., 1993;  Todd et al., 1998).

[11] Generally researchers can run supervised or unsupervised classifications on radiometrically aligned and co-registered multispectral data to separate out forested areas from unforested areas and calculate changes in forest cover by subtracting the aerial coverage of the forested class for each year. This is an entirely separate methodology that is not specifically discussed in this paper but is an additional possibility available to researcher who may want to obtain forest change data. This methodology, however, is also limited in its ability to distinguish spectrally between forests and other highly vegetated areas by the spectral resolution of multispectral data. 

[12] Because the majority of vegetation changes associated with dams are close to 100%, I did not expect to find drastic differences between NDVI and SMA results and therefore did not include these comparisons here.