*2.3. Reference Data*

To reflect the major land-cover classes that are present in the area, reference data were derived, based on our expert knowledge and information, from CORINE, LUCAS, and LPIS land-cover database. Since reference data in the aforementioned databases vary in spatial and semantic consistency, a hybrid classification scheme was devised for this research. Therefore, higher thematic levels from CORINE, LUCAS, and LPIS database were visually interpreted from a time-series of Landsat and Google Earth high-resolution imagery and reduced to the following eight major land-cover classes which were sampled in the study area: cropland, forest, water, built-up, bare soil, grassland, orchard, and vineyard (Table 3). Training and validation pixels were selected at random from the polygon-eroded CORINE and LPIS land-cover maps, and a maximum pixel threshold of 300 pixels per class was set. Afterwards, signatures of the proposed hybrid land-cover classes were checked with LUCAS sample points and visually confirmed from a time-series of high-resolution imagery. This threshold was set following the recommendation from Jensen and Lulla [54] that a number of training pixels should be 10 times the number of the variable used in

the classification model. This hybrid approach was proposed in this research, in order to ensure the reproducibility and optimal number of LC classed were chosen since the difference in the number of distinct classes can affect the classification accuracy [55].

**Table 3.** Description of the major LC classes used in this research, with included codes of CORINE Level 2/3 and LUCAS classification scheme.

