**3. Results**

Resampling the UAV orthoimage to a lower spatial resolution, i.e., a larger value of GSD, resulted in fewer and faster extractions of object boundaries compared to the number of extracted object boundaries generated at the original size of the UAV orthoimage. The processing time for one boundary map was 1–2 min. A larger GSD, at the same scale and merge values, resulted in fewer boundary extractions (Table 2, Figures 4 and 5).

**Table 2.** Ground sample distance (GSD) and number of pixels after image pre-processing.


(**c**)

**Figure 4.** Scale/merge level and number of extracted objects from the resampled UAV orthoimages (**a**) ground sample distance (GSD) 25 cm, (**b**) GSD 50 c, and (**c**) GSD 100 cm. (**a**–**c**) Grey labels—number of extracted objects outside the range, black labels—the lowest number of extracted objects per scale and merge parameter value.

**Figure 5.** (**a**–**i**) Examples of extracted boundary maps. (**a**–**c**) GSD 25 cm; (**d**–**f**) GSD 50 cm, and (**g**–**i**) GSD 100 cm. (**a**,**d**,**g**) Extracted objects at scale 70 and merge 99. (**b**,**e**,**h**) Extracted objects at scale 75 and merge 99. (**c**,**f**,**i**) Extracted objects at scale 80 and merge 99.

A lower scale level and merge level resulted in a higher number of extracted object boundaries for each resampled UAV image. A higher scale and merge level resulted in fewer extracted boundaries (Figure 4). In general, for all resampling, the biggest drop in the number of extracted object boundaries was at scale level values within the range from 70 to 80, and merge level values within the range from 95 to 99 (Figure 4). The incremental value of 1, for merge level 95–99, turned out to be very sensitive in dropping the number of extracted object boundaries (Figure 4a–c).

The optimal scale and merge levels for an automatic boundary delineation were investigated by filtering out the total number of extracted objects with the minimum area of objects from the reference data. The results of this filtering approach are presented in Figure 6. The results showed that for the UAV orthoimages of higher spatial resolutions, namely a GSD of 25 cm, the optimal algorithm values for cadastral mapping resulted in 80 for scaling and from 95 to 99 for merging. In contrast, for the UAV orthoimages having a GSD of 50 cm and a GSD of 100 cm, the common optimal scale level values were 70–80 and merge level 95–98 (Figure 6). Some exceptions were observed for a GSD of 50 cm, where the scale level was 50, 60, and merge level to its maximum. In general, the results showed that the optimal scale and merge level values suitable for cadastral mapping range from 70 to 80 and from 95 to 99, respectively (examples in Figure 5). The optimal scale and merge level values appeared similar as in the investigation of the influence of different GSDs in extracting objects.

**Figure 6.** Comparison in the number of extracted and filtered objects using different scale and merge parameter values, to the number of objects identified in the reference data set.

For further analysis, optimal extracted objects with scale level 80 and merge level 95 for three GSDs of UAV orthoimages were selected (Figure 7a,c,e). The selection was based on common scale and merge levels for three GSDs as well on the highest number of filtered objects per GSD (Figure 6). The filtering approach was additionally applied to the selected optimal extracted objects, i.e., with scale level 80 and merge level 95, to remove objects under the minimum reference object area (Figure 7b,d,f).

**Figure 7.** *Cont*.

**Figure 7.** (**a**,**c**,**e**) Extracted objects at scale level 80 and merge level 95 for (**a**) GSD 25 cm, (**c**) GSD 50 cm, and (**e**) GSD 100 cm. (**b**,**d**,**f**) Filtered objects of scale level 80 and merge level 95 based on minimum object area from the reference data.

A simplification algorithm was applied to both extracted objects and filtered objects. The results showed that if extracted objects are smoothed and smoothed objects are later simplified, the localization of simplified objects is almost equal to that of the extracted ones (Figure 8). The initial tests show that possible shifts in location are possible when a direct implementation of the simplification algorithm to extracted visible objects is used.

(**c**)

**Figure 8.** (**a**) Extracted objects smoothed by making use of the Snakes algorithm. (**b**) Simplification of extracted objects by making use of the Snakes smoothing algorithm and Douglas–Peucker simplification algorithm. (**c**) Extracted objects simplified with Douglas–Peucker algorithm (in black) and compared to object simplifications on (**b**).

The buffer overlay method was used for the accuracy assessment. The accuracy assessment method was applied to the extracted objects, simplified extracted objects, filtered objects and simplified filtered objects. The results show that there is no significant difference in accuracy assessment results when comparing extracted and simplified objects (Table 3, Table 4, and Table 5). At a buffer width of 2 m, a GSD of 50 cm and a GSD of 100 cm provide a higher percentage of correctly extracted objects compared to a GSD of 25 cm. The percentage of correctly extracted objects was 66% for both a GSD of 50 cm and a GSD of 100 cm (Tables 4 and 5). However, the filtering approach contributed to the increased correctness (decreased completeness) and overall quality, for all GSDs. From the filtered objects, the best results for correctness were recorded at a GSD of 50 cm (Figure 9). The percentage indicated the correctness was 77%, while for the completeness it was 67%.


**Table 3.** Accuracy assessment of boundary extractions for a GSD of 25 cm, scale 80, merge 95.

<sup>1</sup> Percentages of simplified boundaries.

**Table 4.** Accuracy assessment of boundary extractions for a GSD of 50 cm, scale 80, merge 95.


<sup>1</sup> Percentages of simplified boundaries.

**Table 5.** Accuracy assessment of boundary extractions for a GSD of 100 cm, scale 80, merge 95.


**Figure 9.** (**a**,**b**) Filtered objects of scale level 80 and merge level 95—simplified, compared with (**a**) cadastral map and (**b**) manually delineated visual object boundaries used as reference data.
