*4.2. Quality Assessment*

Bringing the scale and merge levels to the maximum resulted in some unextracted and fewer visible objects for the whole extent of the image. Although some of the visible objects were left unextracted, the maximum scale and merge level enabled the detection of a group of objects such as a group of tree boundaries, especially at GSDs of 50 cm and 100 cm (Figure 5f,i). In both cases, the balance between completeness and correctness was hard to achieve. This issue was also reported in [26,28]. For this reason, the filtering approach was applied. It was based on the minimum object size as well as on the total number of the objects, both defined based on the reference data. This allowed us to reduce the risk that some of the visible object boundaries remained unextracted as well as over-segmented.

The optimal scale level of 80 and a merge level of 95, were chosen for all three GSDs, to investigate the impact of the same scale and merge level in different resampling. The selection was based on common scale and merge levels per GSD (Figure 6). However, this does not mean that the chosen scale and merge level provided the best object boundary extraction for each of the GSDs. For instance, for small GSDs, the correctness of extracted boundaries is higher at the maximum scale and merge levels (e.g., Figure 5c). For the same scale and merge level, the correctness grows significantly from a GSD of 25 cm to a GSD of 50 cm. The correctness for a GSD of 100 cm was almost equal to the one for a GSD of 50 cm. Considering that more optimal scale and merge levels were applicable for a GSD of 50 cm (Figure 6) and the difference insignificant when compared to the results obtained for a GSD of 100 cm, a GSD of 50 cm appeared to be better in detecting visible boundaries compared to the other two GSDs.

The quantitative method applied for accuracy assessment to automatically extracted objects, filtered objects and to their simplifications, showed that there was no significant difference between extracted objects and simplified objects. This result indicates that the method applied for simplification can be considered appropriate, i.e., the original location of extracted objects was maximally maintained. Although there was no difference in accuracy assessment, the simplification of extracted (or filtered) objects is significant for proper cadastral mapping. Cadastral boundaries usually are defined by straight lines with fewer vertices.

The percentage of suitable extracted boundaries (compared to reference data), for a scale level of 80 and a merge level of 95, resulted in 74% for the assessment of the completeness and 66% for the assessment of the correctness for the extracted object boundaries having a GSD of 50 cm. However, the filtering approach strongly influenced the accuracy assessment. For filtered extractions, the level of completeness was 67%, and the level of correctness was 77%. These results show that the filtering approach increased the correctness of automatically extracted boundaries, and it reaches almost 80% (Table 4). This was due to filtering out small object boundaries from the boundary map. The excluded small objects were mostly present in tree and built-up areas on the UAV orthoimage, i.e., only outlines of group objects were retained (Figure 7c,d). In road extractions, the achieved values for extractions are around 85% for correctness and around 70% for completeness to be of real practical importance [26,34]. Such percentages can hardly be achieved by the workflow developed for automatic delineation of all visible boundaries since the morphology of cadastral boundaries is usually more complex and not all cadastral boundaries are visible, unlike road boundaries.

The accuracy assessment was based on the manually delineated boundaries, which were defined as reference data (Figure 1b). The visible boundaries were manually delineated on the ground truth UAV orthoimage. It is argued that manually delineated boundaries influence the overall results of the accuracy assessment since different human operators might digitize differently [26]. However, in the selected case study, most of the object boundaries were sharp and the presence of white stones at outlines of the agricultural field contributed to the objectivity of manual digitalization. In addition, the real cadastral data could not be used since they did not correspond with the object boundaries on the image (Figure 1a) and it would not have been possible to outline the potential of the ENVI FX. However, the approach of automatic extraction of visible boundaries is case dependent. To reliably avoid the influence of manually digitized reference data, the following studies should consider a case study where the cadastral map is up to date.

## *4.3. Strengths and Limitations of the Automatic Extraction Method Used*

The ENVI FX module handled the full extent of the resampled UAV orthoimages, and no additional image tiling or image conversions were required. ENVI FX provided closed object boundaries directly in vector format, topologically correct polygons. Therefore, no additional image post-processing step, such as vectorization of detected object boundaries, was needed (Figure 5). Thus, the visible object boundaries generated can be directly used for further processing and analysis within geographic information systems (GIS). Additionally, the final output consists of spatial, spectral and textural attributes which are assigned automatically to each extracted object and saved in the attribute table. The vectorized and geo-referenced visible object boundaries, as interpreted in this research, are crucial in cadastral applications especially for the purposes of land plot boundary delineations. Overall, ENVI FX has the potential to automatically delineate visible cadastral boundaries, especially in rural areas.

A comparison of the results regarding the accuracy assessment obtained in this study and the accuracies obtained in the studies [26–28] cannot be done at this time for a number of reasons. First, not all the reviewed feature extraction methods have been applied to UAV imagery. Second, different UAVs may provide different quality of orthoimages. Third, the nature, size, location, and the characteristics of the study objects are far too different. In order to make a reliable comparison on accuracy assessments of different feature extraction methods, first of all, each method has to be studied individually and later tested at the same study area(s). However, the image processing approach of different feature extractions methods may be comparable.

From the reviewed feature extraction methods that have already been applied for detection of visible cadastral boundaries, it can be seen that the MRS method, ESP method, and mean-segmentation method also do not require further image tiling and the final output of the boundary map was in vector format [27,28]. In contrast, vectorization of detected object boundaries was needed for the gPb contour detection method. In addition, it was reported that the method is inapplicable when processing UAV images of more than 1000 pixels in width and height [26]. Similar issues regarding the vectorization of detected object boundaries were reported in [38], where Canny and Sobel edge detection algorithms were used. In order to obtain topologically correct polygons, an additional feature extraction method was used aiming to connect the edges.

ENVI FX allowed some shadow areas in the UAV orthoimage to be extracted as boundaries; however, these do not represent real boundaries in the field. In order to minimize the influence of shadows on feature extraction, it is recommended to capture images in the local time where the solar zenith angle has the smallest possible value. However, the solar zenith angle depends on the geographic location of the study area. Additionally, some other factors such as weather conditions also influence the quality of captured images. To avoid such issues, it is preferable to capture images on a cloudy day without wind. Although ENVI FX has proved to be efficient, one of its limitations is that it is not an open-source tool like mean-shift segmentation, gPb contour detection, Canny, and Sobel, which might be a reason why it is not often used in the land surveyor community. In addition, the extracted objects from the resampled UAV orthoimages were following the pixel borders and further shape simplification was required to make them comparable to spatial units in cadastral applications.

Considering that the morphology of cadastral boundaries is complex [7], compared to physical boundaries, such as boundaries of roads or rivers, delineation of cadastral boundaries cannot be fully automated at this time, and additionally, the verification of the results has to be done with the participation of landowners and other land rights holders. The limitations on extracting only visible object boundaries lie in the fact that not all visible boundaries (land cover boundaries) represent cadastral boundaries (land right boundaries). For instance, when two agricultural cadastral units leased to the same farmer are farmed as one unit, and vice versa. However, visible object boundaries which coincide with the land right boundaries can be automatically detected and used in cadastral applications. In addition, the UAV-based spatial data acquisition is usually affected by special operational regulations that restrict the use of this technology, in particular in urban areas [18].
