**4. Discussion**

#### *4.1. The Developed Workflow*

The developed workflow aimed to provide a solution for UAV-based cadastral mapping using automatic visible boundary extraction with the ENVI FX module (Figure 2). The developed workflow consisted of four steps: (i) image pre-processing, (ii) boundary detection and extraction, (iii) data post-processing, and (iv) accuracy assessment.

The first workflow step includes resampling of a UAV orthoimage. Here, the results of our case study showed that larger GSDs provided faster and fewer extractions of visible object boundaries compared to the original GSD of a UAV orthoimage. For higher spatial resolutions, i.e., smaller GSDs, considering the selected Scale level and Merge level values, the total number of extracted objects was higher.

The second step, which includes object boundary detection and extraction, is dependent on the scale and merge level. The results, presented in Figure 4, showed that lower values of scale and merge levels resulted in a higher number of extracted objects, which led to over-segmentation by reaching thousands of extracted objects. Considering the total number of the reference objects, it is important to note that a scale and a merge level that provide object extractions close to the total number of objects from reference data are important for automatic delineation of visible cadastral boundaries.

The following step, data post-processing, aimed to investigate optimal scale and merge levels and to simplify the extracted objects. The optimal values based on a filtering approach showed that for all tested GSDs in this study, most suitable scale and merge level values for automatic delineation of visible cadastral boundaries were 70–80 and 95–99, respectively. These values can be considered as optimal scale and merge levels for rural areas in general or areas with characteristics similar to the study area of this research. However, to validate the proposed workflow and optimal Scale and Merge levels in areas with different characteristics, such as areas with a larger number of buildings or areas with trees covering parts of boundaries, further experiments are needed. Hence, the scale and merging levels appropriate for cadastral mapping have been determined and this step can be skipped from the workflow step of data post-processing (Figure 2). The use of the Snakes algorithm for smoothing and the Douglas–Peucker algorithm for simplifying has been shown to be very effective (Figure 8a,b). This approach, when combining both smoothing and simplification algorithms, gives better results in terms of a simplified boundary position compared to directly implementing the Douglas–Peucker simplification algorithm, where undesired shifting in boundary position was observed (Figure 8c). In [22], it was reported that the direct implementation of the Douglas–Peucker algorithm was used as a post-processing method in many papers to improve the output by optimizing the shape of objects. However, the simplification approach applied in this study was not examined in the previous papers.

The final step of the workflow was accuracy assessment (see also Section 4.2). The accuracy assessment was based on the buffer overlay method. By increasing the width of the buffer, more extracted boundaries appear to be within the buffer area, which impacts the completeness, correctness, and the overall quality—larger the buffer, the better the results. To have a uniform assessment for all tested GSDs the results were compared at a buffer distance of 2 m. From the reviewed publications presented in Section 1.2, a buffer width of 2 m was also applied in [26–28] as most suitable for the presentation of accuracy assessment results and to avoid uncertainties from resampling effects. However, for comparison to cadastral data, buffer widths should be based on local accuracy requirements [26].

The workflow developed, overall, is in accordance with the general workflow for the cadastral mapping based on suitable computer vision methods for automatic visible boundary extraction provided in [22]. In addition, it provides an additional step and method in data post-processing, such as filtering out irrelevant and small objects from the boundary map, which improves overall quality assessment. Furthermore, it suggests a combined approach for the simplification of extracted object boundaries.
