2.4.2. Hyperspectral Post-Processing

Atmospheric correction is not required when flying at a maximum altitude of 120 m. Once mosaicked, VNIR and SWIR hypercubes are stacked into a single file and the wavelengths associated with water vapour absorbance (i.e., 1350–1460 nm; 1790–1960 nm; 2350–2500 nm [28]) are excluded, resulting in an orthomosaic with 418 exploitable bands (Figure 3(5)).

The Normalized Difference Vegetation Index (NDVI) is used to discriminate the presence of vegetation in the study area (Figure 3(6)) and generate the training set for the final classification step (see Section 2.5). The NDVI map layer is calculated according to Equation (1):

$$\text{NDVI} = (\text{NIR} - \text{RED}) / (\text{NIR} + \text{RED}) \tag{1}$$

with NIR and RED as the near-infrared and red bands centred at 860 nm and 649 nm, respectively. Very low NDVI values (0.1 and below) are associated with bare soil or water, moderate values (~0.3) correspond to shrub and grassland, and high values (0.6–1) are associated with high-density plants and healthy physiological conditions [61]. Since only shrub and grassland vegetation was considered for this work, the NDVI map layer was masked with a threshold of 0.3. The product of this step is a raster with only vegetation distribution.

#### 2.4.3. Endmember Extraction

An endmember is a pure signature for a class [62] and they are essential for the classification of HS data. Pure signatures from hyperspectral imagery can be found using minimum noise fraction (MNF), the Pixel Purity Index (PPI), and endmember extraction techniques (Figure 3). The MNF maximises the noise-to-signal ratio and reduces dimensionality without sacrificing information [63] (Figure 3(7)). Most of the significant information is contained in the first MNF bands, which are used in successive processing, ignoring the rest of the bands containing only noise [64]. The PPI technique [65] searches for pure spectral signatures by identifying the pixels with the fewest mixed spectral signatures (Figure 3(8)). The PPI image locates the pure pixels of the scene, which will then be used to extract the spectra of the potential endmembers [66]. A region of interest (ROI) is a dataset sample considered important for a particular purpose [67]. In this case, the regions of interest (ROIs) contain the pixels with the pure spectra in the scene and are imported in the n-Dimensional Visualizer scatter plot (n-DV; Figure 3(9)). The n-DV is an ENVI tool for visualising the distribution of pixels in the n-D space (where n is the number of bands), allowing the purest pixels representing the spectral endmembers to be identified and clustered (Figure 3(10,11)). After these steps, the classification procedure can be carried out.
