*2.5. Classification*

The classification process uses unsupervised and supervised algorithms to map the spatial location and abundance of each endmember spectrum. The relevant MNF bands are the input for the ISODATA algorithm (Figure 3(12)), which iteratively clusters the pixels using the least distance approach [68]. The result is a first classified image based on the inherent spectral information of the dataset, with each class represented by a different endmember.

The next level of classification is produced by the spectral angle mapper (SAM) algorithm (Figure 3(13)), which calculates the angle between two spectra to identify their spectral similarity based on a maximum angle threshold [69]. This threshold is set to 0.1 radians to minimize spectral mixing issues. This algorithm uses the endmembers of the unsupervised classification to classify.

The last algorithm selected in this study is the support vector machine (SVM) (Figure 3(15)), selected because of the good results it produces with heterogeneous, complex, and noisy data [70]. The SVM separates the classes using a training set with class samples (i.e., support vectors) [71], with every class represented by an ROI. In this case, ROIs are generated based on three sources of information: (1) the unsupervised classification; (2) the SAM-classified image; (3) the NDVI map (Figure 3(14)).
