**1. Introduction**

Establishing a complete land cadastre and keeping it up-to-date is a contemporary challenge for many developing and developed countries, respectively [1,2]. In this research, the distinction between 'developing' and 'developed' countries is considered from a land administration perspective. A developing country refers to a country with low cadastral coverage. A developed country refers to full coverage of a country's territory with defined cadastral land plot boundaries and associated land rights. According to the International Federation of Surveyors (FIG) and the World Bank, only one-quarter of people's land rights across the world are formally recognized by cadastral or other land recording systems [1]. Thus, in developing countries, initial efforts are directed to accelerating cadastral mapping as a basis for defining and recording land rights boundaries and formalizing land-related rights aiming to guarantee land tenure security [3,4]. In developed countries, beyond the initial adjudication stage or establishment of a cadastre, another challenge is the maintenance of person-right-land relation attributes and keeping the cadastral systems up-to-date [5,6]. In countries with a tradition and long history of developing a cadastral system, conventional ground-based cadastral surveying techniques and high positional accuracy of boundary surveying were required. Decades were needed to complete the process of cadastral surveying/mapping and registration [1,6]. Although land cadastres were established, some of the cadastral systems could not be maintained, which led to outdated cadastral maps. Person-right-land relationship is complex and dynamic. Keeping the cadastral system up-to-date

(continuous recording of person-right-land relation attributes, in any land related event, as close as possible to real-time) also requires a flexible and dynamic cadastral system [2,7]. Proposed cadastral surveying techniques are mostly indirect ones rather than ground-based. Ground-based techniques are often argued as being time-consuming and labor intensive [1,5,8].

Emerging tools are mapping techniques based on remote sensing data, in particular, data acquired with sensors on Unmanned Aerial Vehicles (UAVs) [9–18]. Cadastral maps are usually defined as a spatial representation of recorded land plot boundaries or other spatial units that the land rights concern [19]. In general, sensors on UAVs provide low-cost, efficient, and flexible high-resolution spatial data acquisition systems enabling the production of point clouds, Digital Surface Models (DSM) and orthoimages [20,21]. In cadastral applications, UAVs have gained increasing popularity due to the high cadastral mapping potential in a different setting, in rural and urban areas, for developing and developed countries [22]. In addition, UAVs are used for both the creation and updating of cadastral maps [22]. In developing countries, UAV-based cadastral mapping usually serves as a tool for the creation of a formal cadastral system [11–13]. In developed countries, the case studies focus on the assessment of UAVs' data positional accuracy estimation and its conformity with local positional accuracy requirements aiming to use the UAV data for updating existing cadastral maps [14–18]. Here, updating in most cases refers to the comparison of two cadastral maps—one representing the database state, the other recently acquired data. The term updating can be used as a synonym for a "revision" of existing cadastral maps [23]. However, in all case studies reported in [22], cadastral boundaries are manually delineated.

It is argued that a large number of cadastral boundaries are visible and coincide with natural or manmade physical object boundaries [2,24,25]. In the land administration domain, automatic extractions of visible cadastral boundaries have been a recent topic of investigation. The latest studies, though limited in number, assert that visible boundaries, such as hedges, land cover boundaries, etc., which might indicate cadastral boundaries, could be automatically extracted using methods such as algorithms that detect object boundaries in images [22,26–29]. In fact, not all visible cadastral boundaries can be automatically detected—certain boundaries would require a semi-automatic approach, especially in urban areas where the morphology of cadastral boundaries is complex [7]. Nevertheless, the potential of computer vision methods for automatic detection and extraction of visible objects in the images is promising for cadastral applications, especially due to the urgent global need for accelerating and facilitating cadastral mapping as a basis for registration of land rights and following the dynamics of land tenure and land use.

## *1.1. Visible Boundary Detection and Extraction for Cadastral Mapping*

Automatic feature extraction methods from images acquired with high-resolution optical sensors have already proved to be useful for the extraction of boundaries of linear features such as roads and rivers [30–34], and to a much lesser degree, they have also been explored for the purpose of cadastral boundary delineation. A recent study from Crommelinck et al. [22] provides an overview of computer vision methods that might be applicable in the land administration domain for automatic detection and extraction of object boundaries from images acquired with high-resolution optical sensors. Additionally, the general workflow for automatic detection and extraction of visible object boundaries for UAV-based cadastral mapping is provided [22]. The general workflow consists of (i) image pre-processing, (ii) image segmentation, (iii) line extraction, (iv) contour or boundary generation, and (v) image and/or boundary post-processing. Image pre-processing usually includes image conversions, such as resampling or tiling, in order to fit the requirements of a chosen computer vision method. Image segmentation refers to the process of dividing a digital image into non-overlapping objects, which represent homogeneous areas [35]. The third workflow step is the extraction of lines or edges from the segmented images [36]. The next step, contour generation, refers to the extraction of a closed object outlines in the image. In computer vision, they are usually defined as object boundaries, which are derived from connecting edges or lines. An 'object boundary' should encompass an 'object' in an image,

and due to this, both terms are used synonymously in this study. In cadastral applications, objects are usually defined as polygon-based spatial units. The final step, post-processing, includes interventions on the image such as vectorization and/or simplification of automated extraction of objects [26,37]. However, only a limited number of studies have investigated the automatic extraction of objects from images acquired with high-resolution optical sensors for cadastral boundary delineation.

The work by Babawuro and Zou [38] tested Canny and Sobel edge-detection algorithms for the extraction of visible cadastral boundaries from high-resolution satellite imagery (HRSI). In addition, the Hough Transform feature extraction method was used to connect edges and to identify straight lines. The visual presentation of the results showed that the proposed approach can detect agricultural land boundaries, but there were no quantity measures on quality assessment. Kohli et al. [28,29] investigated the use of an object-based approach, namely the multi-resolution segmentation (MRS) and estimation of scale parameter (ESP) to extract visible cadastral boundaries from HRSI. An object-based approach refers to the extraction of object outlines based on a grouping of pixels with similar characteristics and is applied to high-level features which represent shapes in an image [22]. The accuracy assessment in Kohli et al. [28] was pixel-based, and the detection quality in terms of error of commission and omission for MRS were 75% and 38%, respectively. For ESP, the error of commission was 66% and the error of omission 58%. The localization quality for MRS was 71%, whereas it was 73% for ESP, within a 41–200 cm distance from the reference boundaries. Another case of the automatic extraction of visible boundaries based on HRSI is described in Wassie et al. [27]. The study explored the potential of mean-shift segmentation for the extraction of visible cadastral boundaries. The mean-shift segmentation algorithm is a QGIS open source plugin [27]. The object-based measures were applied for the accuracy assessment. Within a buffer distance of 2 m, the percentage indicated the correctness was 34%, while for the completeness it was 83% [27]. The extractions with mean-shift segmentation were closed object boundaries (polygon-based) in vector format and topologically correct. The mean-shift segmentation was applied to a full extent of satellite images. Accordingly, some of the automatic object extraction methods were applied also using UAV images.

The study from Crommelinck et al. [26] outlines the potential of the Global Probability of Boundary (gPb) contour detection method for an automatic boundary delineation based on UAV imagery. gPb is open-source and available as pre-compiled Matlab package. The method was found to be applicable only for processing images of fewer than 1000 x 1000 pixels due to the demanding computing process [26]. The contour map or detected objects were in raster format and required vectorization. Furthermore, Crommelinck et al. [37] discuss the interactive method of visible boundary extractions. The interactive method combines the gPb contour detection, simple linear iterative clustering (SLIC) super pixels and random forest classifier, which allow a semi-automatic approach for the delineation of visible boundaries. The interactive method was tested on visible road outlines based on UAV datasets. The results show that the approach is much more efficient than manual boundary delineation, and all road boundaries were delineated comprehensively.

All the case studies reviewed, both automatic boundary extractions from HRSI and UAV images, have been tested in rural areas since it is argued that most of the cadastral boundaries are visible in such areas [26]. However, not all computer vision automatic feature extraction methods suitable for visible cadastral boundary delineation have already been tested.

Another tool that is also referred to as the 'state-of-the-art' for automatic detection and extraction of features from images is the ENVI feature extraction (FX) module [39,40]. ENVI FX is an object-based module for detecting and extracting multiple object outlines from high-resolution multispectral or panchromatic digital images. The extraction is based on spectral (brightness and color), texture, and spatial characteristics [41]. To the best of the authors' knowledge, there have been no previous publications, nor evidence, that the ENVI FX module has been applied for detecting and extracting visible cadastral boundaries on UAV images.

The justification for using this method is based on Crommelinck et al. [22], in which general workflow and feature extraction methods appropriate for cadastral mapping are provided. The main aim of this study is not to compare automatic feature extraction methods already used for cadastral mapping. Instead, the study focuses on the potential of a feature extraction method which has not been tested yet in cadastral applications. The study can be seen as an important contribution to land administration discussions focusing on cadastral mapping, as there have been a limited number of studies for automatic visible cadastral boundary delineation from imageries acquired using high-resolution optical sensors.
