**3.4 Evaluation of segmentation**

A qualitative evaluation of segmentation output is commonly implemented through visual assessment [31]. This is a rather subjective mean of segmentation accuracy evaluation. Conversely, several supervised and unsupervised approaches

#### **Figure 3.**

*Segmentation results. The boundaries of segments are symbolized with red color (spatial radius: 4, range radius: 160, minimum region size: 700).*

#### *Delineation of Open-Pit Mining Boundaries on Multispectral Imagery DOI: http://dx.doi.org/10.5772/intechopen.94120*

have been presented in order assess image segmentation accuracy. Supervised methods typically compare segmentation output with a reference layer and measure the overlapping area [32]. Unsupervised approaches measure particular features of segments, for example, spectral homogeneity and between object heterogeneity [33]. However, there is not a standard methodology [32].

For the purpose of this study, an unsupervised approach was selected. In specific, the objective function proposed by Espindola et al. [33] was calculated to evaluate the quality of image segmentation results. This function consists of a measure of intrasegment homogeneity and one of intersegment heterogeneity. The first part is intrasegment variance of segments, a weighted average, where the area of each segment represents the corresponding weight. Thus, probable variabilities produced by smaller segments are eliminated. Furthermore, in order to evaluate intersegment heterogeneity, the function employs Moran's I autocorrelation index [34] that measures the spatial association as derived from the total of segments. Moran's I reflects how, on average, mean values of each segment vary from mean values of its adjacent segments. Small values of Moran's I suggest low spatial autocorrelation, hence the adjacent regions are statistically different. This denotes large intersegment heterogeneity. In other words, image segmentation produces segments with discrete boundaries. Employing spatial autocorrelation for evaluating image segmentation quality is especially suitable for region growing algorithms that generate closed polygons [33].

An adequate selection of parameters' values incorporates low intersegment Moran's I index with low intrasegment variance. The proposed function from Espindola [33] adds the normalized values of variance and autocorrelation measures. The objective function and its components were computed for each spectral band of Sentinel-2 imagery. Following, the value of objective function for the entire imagery was calculated by averaging the values of each spectral band. The results are presented in **Table 1**.

As shown in **Table 1**, the mean normalized value of variance slightly changes for different parameters' values and the lowest values corresponds to the lowest values of range radius and minimum region size, as expected. Moran's I index value is decreasing when range radius value and minimum region size are increasing, which means that segments get larger in size but also fewer in number. The selected Mean-Shift parameters' values for this specific study area


#### **Table 1.**

*The values of variance, Moran's I index and objective function for specific Mean-Shift parameters' values.*

(4/160/700) correspond to relatively low Moran's I index value which denotes that neighboring segments are statistically discrete.
