The underlying mechanisms of VIs are well understood, and they emphasise some features of vegetation cover and facilitate obtaining relevant information from digital imagery (Delegido et al., 2013).
In images at ultra-high spatial resolution, it is necessary to determine the VI that enhances the differences among pixels containing vegetation and pixels containing non-vegetation, as well as the threshold value that sets the breakpoint between both classes. The classification output is necessary for the thresholding operation, which needs to be optimised for a successful result.
There are several automatic methods for threshold calculation, among which Otsu’s (Otsu, 1979) method is one of the most utilised for agronomical issues (Guijarro et al.,2011; Meyer and Neto, 2008).
It assumes that the image contains
two classes of pixels (bare soil and vegetation when considering
crop scenarios) and then calculates the optimum threshold based
on minimising combined spread (intra-class variance).
To date, VF has been estimated by relating it to VI values in image
pixels from airborne and satellite platforms, in which the pixels
include vegetated and non-vegetated zones due to the large size
(from a few square metres to square kilometres) (Barati et al.,
2011; Gitelson et al., 2002). Today, the ultra-high resolution of
UAV imagery allows images in which almost every pixel covers
only vegetation or bare soil, with a low proportion of pixels representing
a mixed coverage. Therefore, VF can be calculated as the
percentage of pixels classified as vegetation per unit of ground surface.
This is particularly relevant when working with crops such as
cereals which are sown in narrow crop rows because the surface
distance between such rows is usually not wider than 15–17 cm.
In addition to adequate thresholding and good spatial and temporal
resolution, another important issue in VF mapping is accurate
spatial and temporal consistency.