*2.1. The Conceptual Models of Semantic Land Tenure Relations*

The lists in Table 1 are a number of key models and concepts capturing land tenure relations. Henssen [30] depicts land tenure as institutionalized people-to-land relationships with his "Subject-Right-Object model" [31]. This basic model of land administration has been further modified by for example highlighting the dynamics of land tenure [32]. The "Land Administration Domain Model (LADM)" is to a large degree an extended and more sophisticated model of the basic model, and has become both a conceptual and descriptive standard (ISO 19152). The LADM covers all land tenure-related data components including parties, legal/administrative units, spatial objects, and data on surveying and spatial representation. The LADM can bridge the gap between land policies and information management opportunities and is adaptable to local situations [33,34].




**Table 1.** *Cont.*

Furthermore, the "Continuum of Land Rights" approach emphasizes that land tenure arrangements vary along a continuum of land rights. Not only documented formal land rights are legitimate, but also undocumented informal land rights may exist, and society may accept or condone these alongside formal rights. The continuum of land rights approach is useful in describing de facto land tenure, which is much more fluid and flexible than the static and unchanging (spatio-temporal) description of land rights. It allows more flexibility to define and recognize land tenure based on evidence from the field [35].

The "Fit-For-Purpose land administration" approach mainly focuses on building geospatial data framework of large-scale mapping that can address emerging land tenure issues where no reliable land information exists. This framework highlights following constituent principles in order to not only improve recognition of "value-of information" and maximize "cost-effectiveness", but also decrease "capacity-demanding": (1) general boundaries rather than fixed boundaries; (2) aerial imageries rather than field surveys; (3) accuracy relates to the purpose rather than technical standards and; (4) opportunities for upgrading and improvement [36].

"Responsible land administration" expands the conventional notions of land administration with a normative framework. What is distinctive about this concept is that it takes the following aspects into account: the requirement for any land administration system to ensure the representation of multi-stakeholders in order to foster institutional innovation and inclusion; the incorporation of a broad array of scholarly disciplines into the methodological repertoire, in particular, connecting from technical and information sciences to the social sciences and humanities; the need for a proactive stance in laying the foundation of cutting-edge land administration systems design; connecting the global context to the local and vice versa; the continued need to transfer knowledge into the practice, and vice versa [37]. Technical and operational designs of land information systems can only be innovative if particular societal needs embedded in the design process and in the manner in which land administration is based on shared responsibilities.

#### *2.2. Advancement of EO and AI Applications in Identifying Land Tenure Relations*

One of the significant discussion in EO applications for land tenure relations is to provide the institutional and spatial aspects of cadastral boundaries by identifying relationships between physical objects and visual boundaries based on the notion of cadastral morphology [15,16] and cadastral intelligence [17]. The early experiment demonstrates that over 80% of cadastral boundaries coincide with visible physical objects [15]. In line with a previous endeavor, more tailored object-based workflows using extraction algorithms delineate about 50% of parcel boundaries successfully [16]. Investigating technically transferable workflows is a continuing concern within UAV-based cadastral mapping. For instance, both the *gPb* contour detection method and the ENVI feature extraction

(FX) module has proven accurate results of visible object delineation that coincide with cadastral boundaries at completeness and correctness of up to 80% [11,13]. To extract visible cadastral boundaries within Object-Based Image Analysis (OBIA) environment from High Resolution Satellite Imagery (HRSI), the (semi-)automatic feature extraction methods have been employed and tested in rural areas: mean-shift segmentation with the buffer overlay method [18], and both multi-resolution segmentation (MRS) and estimation of scale parameter (ESP) (only able to automatically extract 47.4%) [17].

In light of state-of-the-art methods in land administration, a deep-learning is becoming highly prominent for the detection of cadastral boundaries [12,19]. Recent evidence suggests that deep fully convolutional networks (FCNs) ensures the high accuracy rather than *gPb* and MRS, with results of 0.79 in precision, 0.37 in recall and 0.50 in F-score [19]. For optimizing image segmentation, one study by [12] not only introduced the interactive boundary delineation workflow, but also examined the better suitability of the deep learning in cadastral mapping with convolutional neural networks (CNNs) by comparing random forest (RF) in machine learning: RF-derived boundary likelihoods (accuracy: 41%, precision: 49%), CNN-derived boundary likelihoods (accuracy: 52%, precision: 76%).

Several attempts make to extract, classify and quantify cadastral boundaries using EO data in association with AI technologies (see Table 2). Along with these varying workflows and its image segmentation techniques that employed, however, there is increasing concern over further investigating deep-learning driven image analysis in land administration including image fusion, image registration, scene classification and retrieval and object detection. For remote-sensing image interpretation, the most applicable deep-learning models in remote sensing are: supervised CNN, recurrent neural network (RNN), unsupervised authencoders (AE), deep belief networks (DBN), and generative adversarial networks (GAN) [38]. Although research on effective use of spatial contextual information in remote sensing for land administration is still in infancy, it can substitute the interpreter to a certain extent (not completely) by delving deeply into AI technologies with computer-vision and deep-learning algorithms.


**Table 2.** Earth Observation (EO) data and Artificial Intelligence (AI) delves deeper into the future of land administration and the advanced techniques substitute to a certain extent the feature and boundary extraction for cadastral mapping. However, a number of critical questions remain about the interpretation of semantic land tenure relations using both EO and AI.

#### **3. Methodological considerations**

#### *3.1. A Di*ffi*cult-to-Access Region: North Korea in the Contexts of Fragile and Conflict-A*ff*ected Areas*

There is increasing concern that remotely obtained data using disruptive technologies in fragile and conflict-affected areas (See Figure 1), where has been named hard-to-reach areas, is more worthwhile in optimal decision-making rather than a limited groundtruthing provided by direct observation (for example, [39–41]). According to [42], some countries such as Somalia, North Korea, and some Caribbean and Pacific island economies do not consistently render an account of internal data owing to conflict, lack of data capacity, or other reasons including quality of sources. An additional encounter with data scarcity and reliability is associated with national security issues in any contexts of fragile and conflict-affected regions worldwide.

**Figure 1.** Contextualization of Area of Interest (left: Fragile States Index 2019 (https://fragilestatesindex.org), right: background information of South and North Korea based on Major Statistics Indicators of North Korea 2019 (https://kosis.kr/bukhan/index/index.do) (devised by authors).

Gathering and establishing reliable information for policy-making in pursuing Korean (re-)unification is more significant than ever during a peace-building process. In this process, re-shaping land governance are a fundamental question focusing on land tenure security, transferability, legitimacy and identity in (re-)unification setting [43,44]. Despite the passive attitude of the North Korean government to disclose information, it is possible to obtain data in a direct or indirect manner, such as [45]: official government reports (e.g., Korean Central News Agency: KCNA and Rodong Sinmun); materials from international organizations dealing with humanitarian aid (e.g., FAO, UNDP, UNFPA, WFP, WHO, UNICEF (For the resources, [46] etc.); information from external agencies or observers in cooperation with local authorities or residents (e.g., Hanns Seidel Foundation for EU-funded project on sustainable forestry in North Korea etc.); data acquired through the joint projects (e.g., the North Korean Ministry of Land and Environment Protection: MoLEP, Swiss Agency of Development and Cooperation: SDC, World Agroforestry Centre's East and Central Asia Office for the Sloping Land Management Program (See further details for the project, [47] etc.).

However, there are still great difficulties in collecting land tenure-related information in North Korea since the government rarely discloses or distributes any policy-related documents, data and statistics. The scope of current research on mapping land tenure relations has been very limited such as restoring historical cadastral maps [48–50] that include both geographical and textual land information. However, most research is still highly dependent on secondary data sources. Many studies using EO data for North Korea have been proposed for monitoring land use and land cover (LULC) over the

past several decades (for example, [51–55]). Furthermore, a number of government institutes and think-tanks have already established different types of thematic maps in North Korea using EO data (e.g., agricultural maps; deforestation maps; land cover maps, etc.). More internationally, a platform called 38 NORTH (https://www.38north.org), provides informed analysis of events in and around North Korea using HRSI, as well as develops the digital atlas that was built in the Google Earth platform. Nevertheless, researchers and policy-makers still have faced with difficulties in incorporating land tenure-related data with EO data due to: (1) levels of accessibility: the limited access to North Korean data; (2) methodological levels: complexity of integrating land tenure attributes with EO data; (3) analytical levels: its lower reliability and validity of acquired information.

#### *3.2. Existing Rules for Defining Land Tenure Relations and LULC classifications*

This section provides an overview of the existing rules for land tenure relations in South and North Korea that can identify the data gap between them. The classification methods of land tenure relations are based on diverse land trajectories: by land ownership, land (use) categories, 3Rs, land characteristics and urban planning facilities (See Table 3).

In South Korea (SK), land is divided into private land, government-owned land including State, province and county land, land owned by corporation or judicial person and land owned by non-judicial person, according to the land ownership trajectories. Contrastingly, the Constitution and the Land Law in North Korea (NK) does not tolerate private transactions with land. The State, collective farms, institutions, enterprises and organizations only govern land and local residents have land use rights (LURs). According to the Constitution (Article 21; 22) and the Civil Law (Article 45; 53) in NK, there is no restrictions on the subject of State ownership and only the State can own land. The cooperative entity refers to the form of collective ownership in which cooperatives assume the ultimate authorities for the land that are restricted by the State.

According to the Act on the Establishment and Management of Spatial data in SK (Article 2), a land category means a type of land that is classified under its primary use, and registered in the cadastral record. Land is currently classified into 28 categories to represent the nature, purpose and status of the land. Meanwhile, in NK the Land Law (Article 7) distinguishes six categories of land use classes: agricultural-purpose land; settlement land; forestry land; industry land; waterstock land; and special-purpose land. However, it is not yet clear whether these land use categories correspond to the zoning system or land category system in SK [56]. Based upon the Civil Law and LADM (focusing on 3Rs), the right type in SK includes co-ownership, servitude, lease, ownership, partitioned ownership, superficies, sectional superficies, tenancy, usufruct, and fishing. The responsibility types include keeping a snow free pavement and cleaning a ditch, and the restriction type includes servitude and servitude partly [57]. For NK, we assume that there is no land use regulation through the restriction of private rights (3Rs), since NK does not recognize private land ownership.

A land characteristics survey investigates land-related data from physical, spatial and socio-economic conditions in SK [58]. The 45 types of land use indicators are basic data for the land classification aforementioned. Moreover, land infrastructure (urban planning facilities) refers to facilities determined by urban management plan among infrastructures. The legal grounds is the National Land Planning and Utilization Act (Article 2) [59]. In NK, land use classifications are different. It follows the land characteristics in the same way as the six types of land prescribed in the Land Law. The diversity of land use appears to be very simple when compared to that in SK, although the Urban Management Law in NK does not explicitly stipulate land infrastructure, it identifies buildings and facilities, which need to be managed. These include residential and factory buildings, water and sewage and heating operation, urban roads and river arrangements, landscaping and urban beautification. Roughly, one can distinct 29 types of land infrastructure elements in NK.

Earth observation (EO) is one of the most essential methods for monitoring the earth's surface and its dynamics at regional to global scales [29]. The term land use defines how a certain portion of the surface is being utilized. In other words, a particular land use label identifies the purpose for which humans exploit the land cover [60,61]. The land cover denotes to the biophysical appearance on the land and determined by the elements of the Earth's (sub) surface. For example, a State park may be used for recreation but have a deciduous forest cover [60,61]. In some countries, a formal/ government LULC classification system exists which can easily describe the actual condition and changes of spatial structures of the land and its attached attributes: the U.S. National Land Cover Database (NLCD, USA [62]), the national Dynamic Land Cover Dataset (DLCD, Australia [63]), the European CORINE land cover (CLC, EU [64]), the Land Cover Map (LCM, Korea), and the National Land Use/Cover Database of China (NLUD-C, China [65]). Although these datasets have been developed with different mapping methodologies and criteria (e.g., variations in the classes and thresholds applied, time of data collection, sensor types, classification techniques, use of in situ data etc.) [65], one can utilize it as basic spatial data to support the design of scientific and efficient policies.

**Table 3.** Existing rules for defining land tenure relations in the context of Korean (re-)unification (functional classifications).


#### *3.3. Adopting a New Methodology: Mixed Methods Design and Information Fusion Approach*

A number of image segmentation techniques and workflows have been developed to detect visible land tenure relations with EO data. Each has its advantages (e.g., automation, coverage, up-to-date, cost-effective etc.) and drawbacks (e.g., technological bias, methodological rationale, lack of social sensing etc.). To especially overcome these constraints, a further advancement with more focus on the idea of 'triangulation' is therefore considered which is particularly associated with methods of investigation and sources of data [66]. A triangulation logic is chosen because land tenure relations are unknown in a given context and monoscopic EO data interpretation without valid inferences would misguide to identify transferrable and applicable proxies. Data integration is at the heart of discerning epistemological assumptions from multiple sources to attain narrative illustration, convergent validation and analytic density [67]. In this regard, this study makes an original contribution to when land tenure and EO data integration occurs, what types of EO data are integrated and how we integrate them.

In aiming to derive informative land tenure relations from EO data, one has to rely on both the characteristics of tenure itself and a number of proxies derived from EO data and spatio-temporal combinations of EO that may capture a particular land tenure characteristic. Our research started by adopting the subjects (e.g., who is the right holder?), rights (e.g., what is included in a certain right?), and objects (what physical extension of a right has?) model of land administration [68]. In parallel with ownership, rights may embrace complex set of rules related to the access, use, develop or transfer [69]. In other words, a household can be associated with a particular land parcel where people can live, own, rent or have the right to use [70]. Thus, the analytical premise has been questioned on the basis of underlying assumptions: (1) is it possible to distinguish collective farmland from State land?; (2) can one see land use rights (LURs)?; (3) is there a use right that can be linked to an individual or group?; (4) are there land transfer rights (LTRs)?; (5) are there land access rights (LARs) and restrictions?

However, methodological difficulties, using EO data in accurately conjoining a household and physical extension of a right over land parcels and measuring the quality of linking information, have existed. There have been only few empirical investigations into decision-making what proxies shall be operationalized based on both theoretical and practical grounds. In terms of using terms 'proxy' rather than similar terminologies such as 'interpretation key', 'index', 'indicator' or 'variable', we follow a definition labelled by [71] that refers to "use of observable physical features or directly measurable variables to understand and extract what actually exists on the ground, but what is not directly observable or measurable from remote sensing data."

Our approach comes from multiple sources, namely EO data, especially focusing on LULC information, *prior* (expert) and contextual knowledge on land tenure relations acquired through previous experiences and perceptions. In view of all that has been mentioned so far, one may suppose that 'information fusion' approach must be considered in order to extract and conciliate significant elements for the semantic (re-)interpretation and, subsequently, for decision-making [72]. Generally accepted disciplines for the notion of information fusion include: psychology, human factors, knowledge representation, artificial intelligence, mathematical logic, and signal processing [73]. It has been noted that transforming data into knowledge is most striking feature of information fusion and must be converted into a certain language or presented by other means such as visualization techniques [73]. Moreover, this method emphasizes that a wide range of structured/unstructured or primary/secondary data sources address semantic relationships and co-occurrence between them [74].

According to [75], an information element is regarded as "an entity composed of a definition set and a content set linked by a functional relationship called informative relation, associated with internal and external contexts". This highlights that one given single data set do not qualify or quantify to make it informative. When answering the research questions or testing relationship between proposed proxies and ground-truth, it is important to quantify recurring spatial attributes and uniformity or distinctiveness in qualitative data allowing rigorous analysis and to determine rational and optimal proxies. It is therefore considered that triangulation logic and information fusion approach would usefully supplement and extend the methodological and epistemological assumptions of semantic land tenure relations through EO data interpretation. Having defined what EO data proxies and information element meant, different types of information element, thus, should be included that enable the EO data proxies to identify land tenure relations logically. Figure 2 depicts the workflow and the main components of the information acquisition and interpretation process of the semantic land tenure relations.

**Figure 2.** The general structure of an information element and its processes for the interpretation of the semantic land tenure relations (devised by authors).
