Article of the Month - April 2025

Artificial Intelligence for Querying Land and Property Data from Cadastral Plans

Hamid Hosseini, Behnam Atazadeh and Abbas Rajabifard, Australia


Hamid Hosseini	Behnam Atazadeh	Abbas Rajabifard

This article in .pdf-format (16 pages)

SUMMARY

Cadastral plans are used in land registration systems for defining legal boundaries of land parcels and properties as well as their associated rights, restrictions, and responsibilities (RRRs). However, existing registered cadastral plans are in 2D non-machine-readable formats and data within these plans are not easily accessible and readily usable, leading to unnecessary delays, disruptions, and costs within land development projects. Artificial intelligence (AI) as an emerging technology has been recognized as one of the operational parameters for advancing land administration systems (LASs) which can offer transformative solutions to overcome traditional approaches. This paper presents a new approach to efficiently retrieve land and property information from cadastral plans, reducing the high cognitive load associated with manual approaches. Our approach’s two core functionalities are data extraction from plans using computer vision and communication with plans using natural language processing (NLP). To demonstrate our approach, a prototype chatbot employing generative pretrained transformer (GPT) as the core large language model (LLM) was developed for data querying from plans. Initial testing shows effective handling of semantic queries, while highlighting the need for further refinement and development in handling more specific queries within land administration domain and complex spatial queries.

1. INTRODUCTION

Effective management of land and property data during its lifecycle is significantly important for the operational efficiency of four critical land administration functions: land ownership, value, use, and development. This results in promoting economic development, environmental sustainability, and social well-being in all jurisdictions and countries (Williamson et al., 2010). Considering land ownership as the basis for the next administrative activities, cadastral plans are common data used for defining and registering boundaries of land parcels and properties as well as their associated rights, restrictions, and responsibilities (RRRs). In Victoria, Australia, plan of subdivision (PS) and plan of consolidation (PC) are currently used to document and represent legal information about the ownership and extent of RRRs over land parcels and properties. In addition to PSs and PCs, abstract of field records (AFRs) are used for documenting and representing the required survey information such as land parcels’ connection to a road intersection or a Crown boundary for generating new plans (Land Use Victoria, 2024b). Any land transaction such as subdivisions, consolidations, and boundary realignments involving new legal boundaries or modifying existing boundaries must be supported by performing land surveying and land registration activities to issue new titles for the new land parcels. These plans are used by relevant stakeholders involved in all stages of land development projects from initial design to future maintenance and form the backbone of various land administration processes such as back-capturing and examination.
However, existing registered cadastral plans are in 2D non-machine-readable formats such as papers and scanned documents which have static representation and lack intelligence. These plans provide essential land and property information. However, due to the special characteristics of these plans, administrative, legal, and survey data within these plans are not easily accessible and readily usable, leading to insufficient data queries. The most apparent characteristics of these plans are (see Figure 1):

Dense and detailed data: Each sheet of the plans contains dense and detailed textual elements (e.g., characters, numbers, and punctuations) in multiple sizes and orientations as well as geometric elements (e.g., symbols, lines, polygons). Although it provides rich and comprehensive understanding, it leads to difficulties in finding specific land and property information quickly.
Fragmentation: Data inside these plans are fragmented and scattered not only across multiple sheets but also throughout individual sheets. Although it directs attention to a specific portion of the plan’s information, it makes it difficult to follow the content coherently and may cause to miss critical land and property information.
2D flat view: Elevation and depth information which is essential for representing the vertical dimension is provided by 2D flat viewed diagrams such as crosse-sectional and isometric diagrams, leading to ambiguity in terms of visualisation.
Isolation: Considering two or more land parcels and properties, their own plans are stored separately, leading to difficulties to find out their relationships in an integrated

Figure 1. A Crown plan of a tunnel in the city of Melbourne with its characteristics in static PDF format, leading to increasing much cognitive load to understand the content

Overall, a much cognitive load is required to understand the content inside cadastral plans and therefore query desired data, especially for less experienced users with less familiarity with the content, leading to a reduction in the level of accessibility and usability of these plans. This can potentially result in slow and inefficient administrative processes which can lead to unnecessary delays, disruptions, and costs within land development projects, particularly in large-scale infrastructure projects which deal with numerous land parcels and properties.

Transforming data from 2D non-machine-readable formats to 3D full digital models based on Land Administration Domain Model (LADM), City Geography Markup Language (CityGML), and Industry Foundation Classes (IFC), few studies have been conducted for connecting LADM and IFC (Atazadeh et al., 2018), integration of LADM and CityGML (Góźdź et al., 2014; Li et al., 2016), and 3D-extending the CityGML for underground legal boundaries (Saeidian et al., 2024). However, the majority of existing registered land and property data have yet to be mapped into these newly developed 3D digital models. In Victoria, Australia, although many plans have converted to digital records (i.e., LandXML) within the back-capturing process under the Digital Cadastre Modernisation program (Land Use Victoria, 2024a), this initiative currently does not support multi-story properties (Cumerford, 2010). Considering the process of mapping, it is a semi-automated task requiring domain experts and is not usable by non-specialist stakeholders. In addition, conducting new surveys is both time-consuming and costly, and using crowdsourcing approaches may fail in terms of accuracy and heterogeneity.

Applying innovative and efficient ways to enhance the reusability of the existing data can potentially benefit stakeholders such as land surveyors and land registries, facilitating access to and retrieval of land and property information for different purposes. The new intelligent approaches should be able to review the plans quickly and generate query results faster, assisting stakeholders to make smarter decisions with reduced cognitive effort. Artificial intelligence (AI) has been recognized as one of the operational parameters for advancing land administration systems (LASs) (Chehrehbargh et al., 2024). AI has been widely adopted in various domains such as geospatial science and has resulted in the emergence of geospatial AI (GeoAI). The adoption of AI into land administration, as a subdomain of geospatial science, can offer intelligent solutions to overcome traditional approaches within land administration practices. By developing AI models, it becomes possible not only to review and query from plans directly but also to accelerate their conversion to 3D digital models and hence effective data validation, storage, visualisation, and query. This is in line with the future visions defined in the Cadastre 2034 initiative which has a vision to enable people to understand their RRRs related to land and real property in a survey accurate and 3D environment. The aim of this initiative is to achieve a cadastral system which is sustainably managed, truly accessible, easily visualised, readily used, fully integrated with broader interests on land and provides a dynamic, 3D digital representation of real world (ICSM, 2019).
The main purpose of this paper is to introduce a new AI-based approach to support land administration stakeholders in querying land and property data from existing registered cadastral plans in an intelligent environment using computer vision and natural language processing (NLP). This is expected to increase the efficiency of the document reviewing process and assist the stakeholders to conduct land administration tasks with considerably less cognitive load. To demonstrate the practical applicability of our approach, an initial prototype of a chatbot has been developed and tested in which users can upload a cadastral plan in PDF format and ask questions about the plan in a natural language form and receive a response accordingly. This serves as a proof of concept, illustrating how AI can transform traditional land administration processes, making cadastral plans more accessible and reusable.

The rest of the paper is organised as follows: Section 2 provides the background relevant to the research. In Section 3, the proposed approach is described. In Section 4, an initial prototyping and testing has been conducted to examine the feasibility of the proposed AI-based approach for querying land and property data. Finally, Section 5 provides discussions and conclusions.

2 BACKGROUND AND RELATED WORK

2.1 Artificial Intelligence

Although AI lacks a universally accepted definition, it is generally recognized as enabling machines to replicate different aspects of human intelligence such as reasoning, learning, perceiving, communicating, problem-solving, and acting (Russell & Norvig, 2016). This can lead to learning from experience, adapting to new situations, and performing human-like tasks (Duan et al., 2019). It involves a wide range of techniques that can broadly be categorized into rule-based and data-driven paradigms. Reasoning aspect of human intelligence refers to the process of human-like logical thinking. Expert systems are popular example of AI developments related to reasoning aspect in which explicit knowledge in form of encoded if-then rules is used (Gupta & Nagpal, 2020). As computational power grew, this rigid approach has been replaced by data-driven approaches which are based on learning from data. Although it has less transparency, it brings more flexibility and adaptability to new situations. Machine learning algorithms such as decision trees, random forest (RF), support vector machine (SVM), and k-means can be fed with experienced data and be trained and make predictions for new data (Zhou, 2021). More advanced, deep learning as a subset of machine learning can extract deep patterns from data through its multi-layer neural network (Goodfellow, 2016). Moreover, computer vision and natural language processing (NLP) can replicate human cognition aspects such as vision and speech. Computer vision can perceive and understand visual information and has shown its capabilities for information extraction from imagery data such as plans. On the other hand, NLP can potentially understand and generate information in for of human language such as textual data (Nishant et al., 2020). Leveraging these techniques, AI offers three main capabilities: 1) automation 2) real-time functionality and prediction 3) intelligent decision-making.

Optical character recognition (OCR) is a technology used to identify and convert textual data form different types of documents, such as scanned papers and PDFs into machine-readable format (Memon et al., 2020). OCR for survey plan analysis automates the extraction of textual information, such as boundary descriptions, parcel numbers, and surveyor annotations, from scanned survey documents. Traditional OCR used pattern recognition techniques and rule-based approaches but data-driven OCR like those powered by convolutional neural networks (CNNs) can handle more complex tasks, such as identifying text in varied fonts, layouts, and even handwriting. These systems learn from large datasets and continuously improve their accuracy and efficiency through AI models. By converting survey plans into machine-readable textual formats, OCR streamlines land tenure documentation. This automation enhances the efficiency of land administration tasks, reducing manual data entry and errors.

Upon extraction of textual information from plans, these raw texts need to be processed to be converted into computer-intelligible (i.e., numerical) format (Chen et al., 2022). Moreover, for having enhanced communication with the plans, generating new textual data is required. These can be done using NLP and large language models (LLM). NLP consists of preprocessing tasks for cleaning the text and vectorization of words for converting the text into numerical format. Preprocessing consists of several steps as follows:

Removing unnecessary data: It refers to removing punctuation, HTML tags, etc.
Tokenization: It refers to splitting text into smaller units such as words or sentences.
Normalization: It refers to standardizing text by converting all characters to lowercase and normalizing spelling.
Removing stopwords: It refers to removing words that do not contribute much meaning (e.g., the, and, or).
Stemming and lemmatization: Stemming refers to reducing words to their base or root form and lemmatization refers to mapping words to their dictionary form.

In order to convert the prepossessed text to machine- readable format, the text needs to be converted into numerical format (i.e., vector representation), known as embedding vectors. Bag of Words (BoG) (Rani et al., 2022) is a traditional model for word embedding which relies on the frequency of word occurrences but lacks contextual awareness, resulting in semantic inaccuracy. In contrast, Word2Vec (Mikolov et al., 2013) and GloVe (Pennington et al., 2014) are predictive models based on learning concept. GloVe captures global context across the entire corpus and is more suitable for representing semantic relationships. However, these models are static and assign a single vector representation to a word regardless of its context. On the other hand, transformer-based word embedding methods such as embeddings from language models (ELMo) (Peters et al., 2018), bidirectional encoder representations from transformers (BERT) (Devlin, 2018), and generative pretrained transformer (GPT) (Brown, 2020) excels in contextual embedding in which dynamic vectors as the output adapting to the context are generated. These methods are based on transformer architectures which use self-attention mechanism to achieve a deeper understanding of contextual relationships (Vaswani, 2017). Self-attention mechanism allows each word to attend to all other words in the sequence and generate a set of context-sensitive vectors for each word, enriched with weighted information from other words in the sequence. The relevance of tokens to one another is quantified through attention scores which is calculated by using the dot product of query and key vectors, followed by a softmax operation for normalisation. Considering LLM models for generating new textual data related to a given text, the embeddings are updated and refined layer by layer within the transformer. Once the sequence has been processed, the new textual data is generated by selecting the next word based on the highest probability from the softmax output. This new textual data may include summaries or answers to specific queries about the input text.

2.2 Related Work

Several studies have been conducted for information extraction from cadastral document. In (Lenc et al., 2021), fully convolutional networks (FCNs) has been developed for landmark and border line detection combined with traditional image processing techniques like edge detection for facilitating the creation of maps from historical documents. Applicability of neural networks for effective annotation in historical maps to facilitate their automatic vectorization is discussed in (Petitpierre & Guhennec, 2023). In (Lenc et al., 2023), integration of neural networks with standard computer vision techniques for the automatic analysis of historical cadastral maps has been suggested when little training data are available. In (Mango et al., 2023), line convolution neural network (LCNN) and ResNet-50 have been used for detecting parcels and their numbers in paper-based cadastral data, respectively. However, it is unable to detect all numbers. In a similar study (Marcial et al., 2013), lot numbers have been recognized using artificial neural network (ANN) and image processing techniques like binarization. The study achieved an average detection rate of 90% for smaller maps and 84.78% for larger maps.
In (Franken et al., 2021), a data processing platform named VeCToR has been developed that combines deep learning algorithms with human validation for high accuracy in the extraction of geometric and semantic information from millions of historical field sketches which are schematic drawings and different from cadastral plans, aiming to rebuild cadastre maps of Netherland. Conversion of AFRs files into LandXML files using OCR process has been investigated in (La Rosa & Garrido, 2019). In (Yıldız et al., 2021), a model has been developed to automatically digitize the temporal dimension of cadastral parcels using OCR and EAST DL text detectors. However, it has challenges with ambiguous texts, light reflections, and blurry images. Overall, deep learning models can achieve excellent performance in cadastral map digitization, but the limited training data is a big challenge, especially for historical maps (Ignjatić et al., 2018).

3. PROPOSED AI-BASED APPROACH

Considering the issues associated with 2D survey plans and the transformative solutions that AI brings to us, a conceptual framework with an architecture has been introduced which is illustrated in Figure 2. The flow begins with user interactions, where users can upload survey plans in raster PDF formats. Then they can input queries seeking specific information such as depth limitation or distance between two specific points ensuring accessibility and ease of communication, regardless of their technical background. Based on the uploaded plans and input queries, users can receive a coherent response in natural language. This framework is underpinned by two core functionalities: data extraction from plans using computer vision and communication with plans using NLP, which are described in the following subsections.

3.1 Textual and geometric data extraction from plans

Through this functionality, the uploaded plans undergo conversion to an image format such as PNG to be prepared for extracting essential data from them. Once the sheets of the plan are converted to an image format, key components within each image, such as notations and boundaries, need to be identified and segmented using computer vision techniques. In this regard, an essential component of the process is the utilization of CNNs, which excel in image segmentation tasks by providing a precise and detailed examination of the details present in survey plans. These networks perform a thorough and detailed analysis of the various elements present in the survey plans, ensuring that all textual and geometric components are captured and ready for further analysis. The output is a segmented image in where each pixel is classified into distinct categories such as annotation, boundary, and other relevant features. Following the segmentation process, textual data within the segmented image regions containing texts are extracted and transcribed from image format into machine-readable text using OCR engines. This results in raw textual output being stored in a database or in a file format like LandXML. Simultaneously, geometric data from the survey plans are extracted and converted into structured formats like Geo JavaScript Object Notation (GeoJSON) which leads to representing and storing spatial data. Overall, this functionality ensures that both textual and geometric information are systematically extracted, stored, and made readily available for further analysis.

3.2 Textual query processing and response generation for users

Through this functionality, interaction between users and plans is enhanced. User queries which are in natural language need to be analysed and interpreted in a way that allows machines to understand. In this regard, a fine-tuned LLM based on pretrained LLMs such as GPT must be applied to handle the queries. In fact, the NLP component bridges the gap between the extracted textual and geometric data and the natural language queries submitted by surveyors. LLM, specifically trained on a purpose-built dataset of surveyor queries related to survey plans and corresponding answers, can potentially accommodate various queries. The queries can be a straightforward semantic query that its corresponding answer has been explicitly stated within the plan (e.g., what is the reduced level of point no. 123?) or more complex spatial query (e.g., what is the distance between point no. 123 and no. 124?).

First, the query is classified to find out whether it is semantic or spatial before accessing the structured database. This classification is conducted using supervised learning algorithms. After classification, the query is converted to a structured format like JSON to facilitate the retrieval of information that has been extracted and stored in a database before. Upon receiving results from the database, as these outputs are often presented in a structured format and may not be familiar for the user, the information is converted back into a natural language format using the fine-tuned LLM, providing a description of the result. Moreover, if the query contains both semantic and spatial components, the query is sent to a developed spatial reasoning model to perform spatial calculations. This component primarily uses symbolic reasoning such as if-then rules and algorithms (e.g., topological operators, metric operators, and directional operators) to process structured geometric data and it does not use data-driven approaches. The output might be precise spatial measurements or analysis results that can be integrated into the final response provided to the user. Finally, the user receives relevant responses.

Figure 2. The proposed architecture for design of the prototype

4. INITIAL PROTOTYPING AND TESTING

To prove the concept, an initial prototyping was conducted to test the feasibility of data querying from plans that include complex spatial layouts and semantic annotations using AI technologies. The prototype took the form of a web-based chatbot developed using Python programming language in which GPT-3.5-Turbo was employed as the core LLM. To have a customized LLM, we utilized retrieval-augmented generation (RAG) techniques, which combine generative capabilities with the ability to access external information sources, significantly improving the accuracy and relevance of responses in land administration scenarios. In this method, the output of a pretrained LLM is optimized by referencing an external knowledge base outside of the LLM training data sources before generating a response. First, the embeddings vectors of the stored chunks (i.e., paragraphs) are first generated in a pre-trained LLM. The embedding vector of user’s prompt is first generated and the similarity between the user’s prompt and the stored paragraphs’ embedding vector is then calculated to retrieve the most relevant contexts. We used cosine similarity method and Oracle database containing 1901 paragraphs extracted from scientific papers in different sources such as Land Use Policy journal, organizational reports such as FIG and CSDILA reports and papers, and governmental publications such as resources in Land Use Victoria and Victorian legislation. The most relevant contexts are then combined with user’s prompt and a comprehensive prompt is formed as an input for the pretrained LLM. Responses will be generated tailored to land administration domain that enhances the responses of the pretrained LLM and reduces the occurrence of hallucinations, thereby increasing the models’ credibility. The interface of the developed prototype is depicted in Figure 3. It includes a drag-and-drop file upload feature that allows users to upload survey plans in PDF format. Also, a textbox is considered that lets users type questions related to the uploaded plan and then click the submit button to process the query, which triggers the system to extract and interpret relevant information from the plan and generates suitable response, which is then displayed the chatbot response textbox, providing users with the requested information in real time.

Figure 3. The interface of the developed prototype

To evaluate the prototype, it was then tasked with answering targeted questions to test its ability to understand semantic content and spatial relationships within survey plans. The performance is assessed based on the model’s accuracy/correctness in identifying and interpreting data such as boundaries, annotations, and any semantic notes within the survey plans. The testing phase used various semantic and spatial queries derived from the Crown plan depicted earlier in Figure 1 as a case study. The prototype’s responses were then evaluated and compared against the expected results. Several queries were executed on a plan, with specific portions utilized for testing, as illustrated in Figure 4. The detailed queries and corresponding responses generated by the prototype are shown in Tables 1 and 2. This evaluation highlights the steps necessary to achieve greater robustness in real-world applications in processing survey data queries.

Figure 4. Some parts of a Crown plan as initial data for testing and evaluation

Table 1. Semantic queries and corresponding responses

Query	Response	Accuracy score
*Is this a plan of subdivision or a plan of consolidation?*	This document is a plan of subdivision. The plan shows the division of Crown Allotments into smaller parcels. The legend indicates "Crown Allotment" and details of the subdivision of those allotments.	0
What is the plan number? What is the edition of plan? What is the number of the crown allotment?	The plan number is OP 119954. The edition of the plan is 3. The crown allotment number is Q2.	100
*What is its city, parish, and country?*	The plan shows the City of Melbourne and the City of Richmond, Parishes of Melbourne South, Melbourne North, and Jika Jika, all located in the County of Bourke. The country is not explicitly stated on the plan, but given the other details, it's implied to be Australia.	100
*Which datum is used for reduced level?*	The Australian Height Datum is used for the reduced levels.	100
*Who is the Licensed Surveyor?*	Neil W Street of 437 St Kilda Road, Melbourne is the Licensed Surveyor.	100
*Who is the Surveyor-General?*	P. Williams is the Surveyor-General	0
*Who is the examiner?*	P. Williams examined the plan.	100
*On what date was it signed, and by whom?*	The plan was signed on July 3, 2002, by Neil W. Street.	50
*What is the scale of the map?*	The main plan scale is 1:3000. There are also details at scales of 1:500, 1:1000, and 1:750 and 1:10. One vertical section is at 1:5000 and another is at 1:5500.	100
*What does CH mean in the legend abbreviation?*	CH stands for Chainage along the Highway Centre Line.	100
*How is the parcel shown?*	The parcel is shown by continuous thick lines.	100
*What is the CAD file number of sheet 2?*	The CAD file number for Sheet 2 is SGLA5228.	100
*How many sheets does it have?*	The plan consists of 19 sheets. There is also a modification table as a 20th sheet.	100
*How many abbreviations are there in the legend?*	There are eight abbreviations in the legend.	0

Table 2. Spatial queries and corresponding responses

Query	Response	Accuracy score
What are the coordinates of point numbers 17 and 19?	Point 17: E 321761.289, N 5811548.400, RL -16.601 Point 19: E 321775.778, N 5811548.505, RL -38.372	50 (Do not have RL)
What is the distance between point numbers 17 and 19?	I cannot calculate the precise distance between points 17 and 19. While the Easting and Northing coordinates are provided, I lack the information necessary to perform a coordinate transformation to accurately compute the distance in meters. This would require knowing the map projection used (e.g., UTM, State Plane) and potentially applying a more complex calculation than a simple Euclidean distance. To obtain the distance, you would need to use a Geographic Information System (GIS) or specialized software that incorporates the necessary projection information.	Can not
Which points have the vertical dimension within their coordinates?	The points listed in the coordinate table on Sheet 18 and Sheet 19 include a "RL" value alongside their Easting and Northing coordinates. "RL" stands for Reduced Level, which represents the vertical dimension (elevation). Therefore, all the points listed in those tables (points 1 through 172) have a vertical dimension included in their coordinate information.	0
What is the RL of point number 1?	The RL of point number 1 is -5.398.	0
Tell me which point have RL?	I cannot definitively identify which points have RL values without a more thorough and systematic review of the tables on sheets 18 and 19. The OCR results are not perfectly clean and may have missed some RL values, or they may be inconsistently formatted. A manual inspection of the original scanned document is required to confirm which points include RL data.	Can not
Is point number 59 a VC or TP?	Point number 59 is a VC (Vertical Curve Tangent Point).	100
Which type of map scale is used? Linear or ratio?	The map uses a ratio scale. A ratio scale (also called a representative fraction) expresses the scale as a ratio, such as 1:3000, meaning one unit on the map represents 3000 units on the ground. This is different from a linear scale which uses a visual bar to show the scale. While the document includes a linear representation of the scale, the fundamental scale used is expressed as a ratio.	50 (It is not correct for the scale on the right side of the sheets)
What does DETAIL 4 show?	DETAIL 4 shows a cross-section view of a portion of the highway and surrounding area, focusing on the area around point numbers 113-117. It highlights the relationship between the highway, the natural surface level, and other details of the subsurface. The detail drawing provides measurements and levels relevant to this specific part of the project.	0
Is DETAIL 4 a horizontal map or vertical?	DETAIL 4 is a vertical cross-section. It shows a slice through the earth, revealing the vertical relationships between the highway, the ground surface, and other features at that location. It's not a horizontal plan view which would show the spatial layout from above.	0
Is SECTION A/1 a horizontal map or vertical?	SECTION A/1 is a vertical cross-section. The drawings labelled "SECTION" in this document are all vertical cross-sections showing subsurface details and elevations, not horizontal plan views.	100
How many points are observed in DETAIL 3?	DETAIL 3 shows seven numbered points: 153, 154, 155, 157, 161, 163, and 165.	0
How many parcels are there?	Crown Allotment Q2 Crown Allotments 5 & 6, Section 19D Crown Allotment 5C Crown Allotment 1A, Section 39A	0
Is there any easement within this plan?	The provided document does not explicitly mention easements. While the plans show various features and boundaries, there is no text or notation that clearly indicates the presence of easements. To determine whether easements exist, you would need to consult additional documentation related to the land title or other legal records associated with these parcels.	Can not
Where is the boundary of parcel Q2?	The boundary of parcel Q2 is shown as a thick continuous line on the provided maps, encompassing an area that extends along the Yarra River. The exact coordinates of the boundary are not explicitly listed, but the boundary is defined by the series of points and curves shown on sheets 1, 2, 3, 4 and 5. The boundary follows both sides of the Yarra River and is impacted by the highway (as the parcel is a stratum of crown land).	0 (Only in sheet 1)
With which points does Crown Allotment Q2 intersect? Please list the points that define its perimeter.	I cannot provide a complete list of points defining Parcel Q2's perimeter. The maps are complex, and a precise identification requires a manual inspection of the drawings themselves. The OCR data provides the coordinates of many points but does not explicitly connect those points to Parcel Q2's boundary in every case. To create the complete list, one would need to manually trace the boundary line on the maps and note the numbers of the points it passes through.	Can not

5. DISCUSSION AND CONCLUSION

In this paper, a novel method based on AI as an emerging technology has been proposed for retrieving information from cadastral plans, aiming at overcoming limitations associated with the plans, such as fragmentation, and hence decreasing high cognitive load associated with traditional manual analysis. Our proposed framework addresses the challenges by leveraging cutting-edge AI technologies, specifically computer vision techniques for image processing, OCR for text extraction from the plans, and NLP and LLMs to facilitate query processing and response generation. This approach significantly streamlines the data access and allows quick retrieval of information from cadastral plans.

Results from the initial testing phase show that the developed prototype can effectively handle semantic queries explicitly defined within the plans. However, more extensive fine-tuning is needed to enhance LLMs’ capabilities in handling land administration domain-specific queries. Moreover, the performance of the prototype decreased when dealing with spatial queries. While it could extract spatial coordinates of points, it cannot perform the necessary spatial calculations or contextual understanding to provide meaningful spatial insights. This indicates a need for employing spatial analysis tools to provide the chatbot with accurate spatial reasoning. In summary, the prototype demonstrates potential for automating document analysis, especially for simple fact-extraction tasks. However, improvements are needed to enhance the chatbot’s ability to understand implicit information and infer relationships between different parts of the document which can potentially lead to more accurate and complete responses.

Regarding improvement of the current state, high-quality data is required for training algorithms to make accurate predictions within different components of the prototype which is a critical area for future research. High-quality data leads to ensure the developed AI models can recognise and interpret various textual and geometric elements within these plans, such as legal boundaries, survey observations, and administrative information. Moreover, the performance of the models in handling new situations in different survey use cases is significantly related to the diversity of the datasets and various land administration concepts such as legal boundaries and spaces, survey measurements, and land administration terminologies must be clarified. Hence, standardized data and processing protocols are required to create larger, more consistent, and comprehensively annotated datasets for model training. Additionally, to address data privacy concerns, we propose to use on-device LLMs, such as Llama, instead of cloud-based alternatives like GPT. By addressing these challenges and further developing the prototype, the efficiency and effectiveness of survey information retrieval can be dramatically improved, leading to enhance the robustness of AI-driven solutions for land administration purposes.

REFERENCES:

Atazadeh, B., Rajabifard, A., & Kalantari, M. (2018). Connecting LADM and IFC standards–pathways towards an integrated legal-physical model. The 7th International FIG Workshop on the Land Administration Domain Model, Zagreb, Croatia.
Brown, T. B. (2020). Language models are few-shot learners. arXiv preprint arXiv:2005.14165.
Chehrehbargh, F. J., Rajabifard, A., Atazadeh, B., & Steudler, D. (2024). Identifying global parameters for advancing Land Administration Systems. Land use policy, 136, 106973.
Chen, X., Xie, H., & Tao, X. (2022). Vision, status, and research topics of Natural Language Processing. In (Vol. 1, pp. 100001): Elsevier.
Cumerford, N. (2010). The ICSM ePlan protocol, its development, evolution and implementation. FIG congress, Sydney, Australia.
Devlin, J. (2018). Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805.
Duan, Y., Edwards, J. S., & Dwivedi, Y. K. (2019). Artificial intelligence for decision making in the era of Big Data–evolution, challenges and research agenda. International journal of information management, 48, 63-71.
Franken, J., Florijn, W., Hoekstra, M., & Hagemans, E. (2021). Rebuilding the cadastral map of The Netherlands, the artificial intelligence solution. FIG working week, Amsterdam, Netherlands
Goodfellow, I. (2016). Deep learning. In: MIT press.
Góźdź, K., Pachelski, W., van Oosterom, P., & Coors, V. (2014). The possibilities of using CityGML for 3D representation of buildings in the cadastre. Proceedings of the 4th International Workshop on 3D Cadastres Workshop, Dubai, United Arab Emirates.
Gupta, I., & Nagpal, G. (2020). Artificial intelligence and expert systems. Mercury Learning and Information.
ICSM. (2019). Cadastre 2034-Powering Land & Real Property. https://www.icsm.gov.au/sites/default/files/Cadastre2034_0.pdf
Ignjatić, J., Nikolić, B., & Rikalović, A. (2018). Deep learning for historical cadastral maps digitization: Overview, challenges and potential. Comput. Sci. Res. Notes 2803, 42-47.
La Rosa, D., & Garrido, O. (2019). Conversion of Cadastral Survey Information into LandXML Files using Machine Learning University of Southern Queensland].
Land Use Victoria. (2024a). Digital Cadastre Modernisation. https://www.land.vic.gov.au/surveying/projects-and-initiatives/digital-cadastre-modernisation
Land Use Victoria. (2024b). Plans of subdivision and consolidation. https://www.land.vic.gov.au/land-registration/for-professionals/plans-of-subdivision-and-consolidation
Lenc, L., Baloun, J., Martínek, J., & Král, P. (2023). Towards Historical Map Analysis Using Deep Learning Techniques. 19th IFIP International Conference on Artificial Intelligence Applications and Innovations, León, Spain.
Lenc, L., Prantl, M., Martínek, J., & Král, P. (2021). Border detection for seamless connection of historical cadastral maps. Document Analysis and Recognition–ICDAR 2021 Workshops: Lausanne, Switzerland, September 5–10, 2021, Proceedings, Part I 16,
Li, L., Wu, J., Zhu, H., Duan, X., & Luo, F. (2016). 3D modeling of the ownership structure of condominium units. Computers, environment and urban systems, 59, 50-63.
Mango, J., Wang, M., Mu, S., Zhang, D., Ngondo, J., Valerian-Peter, R., Claramunt, C., & Li, X. (2023). Transform paper-based cadastral data into digital systems using GIS and end-to-end deep learning techniques. International Journal of Geographical Information Science, 37(5), 1099-1127.
Marcial, D. E., Dy, E. D., Maceren, S. F., & Sarno, E. R. (2013). Artificial neural network-based lot number recognition for cadastral map. Recent Progress in Data Engineering and Internet Technology: Volume 1,
Memon, J., Sami, M., Khan, R. A., & Uddin, M. (2020). Handwritten optical character recognition (OCR): A comprehensive systematic literature review (SLR). IEEE access, 8, 142642-142668.
Mikolov, T., Sutskever, I., Chen, K., Corrado, G. S., & Dean, J. (2013). Distributed representations of words and phrases and their compositionality. Advances in neural information processing systems, 26.
Nishant, R., Kennedy, M., & Corbett, J. (2020). Artificial intelligence for sustainability: Challenges, opportunities, and a research agenda. International journal of information management, 53, 102104.
Pennington, J., Socher, R., & Manning, C. D. (2014). Glove: Global vectors for word representation. Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), Doha, Qatar.
Peters, M. E., Neumann, M., Zettlemoyer, L., & Yih, W.-t. (2018). Dissecting contextual word embeddings: Architecture and representation. arXiv preprint arXiv:1808.08949.
Petitpierre, R., & Guhennec, P. (2023). Effective annotation for the automatic vectorization of cadastral maps. Digital Scholarship in the Humanities, 38(3), 1227-1237.
Rani, D., Kumar, R., & Chauhan, N. (2022). Study and comparision of vectorization techniques used in text classification. 2022 13th International Conference on Computing Communication and Networking Technologies (ICCCNT),
Russell, S. J., & Norvig, P. (2016). Artificial intelligence: a modern approach. Pearson.
Saeidian, B., Rajabifard, A., Atazadeh, B., & Kalantari, M. (2024). Managing underground legal boundaries in 3D-extending the CityGML standard. Underground Space, 14, 239-262.
Vaswani, A. (2017). Attention is all you need. Advances in neural information processing systems, Long Beach, CA, USA.
Williamson, I., Enemark, S., Wallace, J., & Rajabifard, A. (2010). Land administration for sustainable development. Citeseer.
Yıldız, F. B., Ayazlı, İ. E., & Takcı, H. (2021). Generating temporal cadastral parcels with artificial intelligence algorithms within the scope of cadastre 2034. Intercontinental Geoinformation Days, 2, 96-99.
Zhou, Z.-H. (2021). Machine learning. Springer nature.

BIOGRAPHICAL NOTES

Hamid Hosseini is a Ph.D. candidate in Geomatics at the University of Melbourne. He is an active research member of the Centre for Spatial Data Infrastructures and Land Administration (CSDILA). Hamid’s research focuses on AI-driven land administration. His research interests involve GIS, BIM, GeoAI, indoor positioning, 3D Land Administration, and Oracle Database Administration, with four years of professional experience in the industry.

Behnam Atazadeh is an ARC DECRA Fellow (Senior Research Fellow) in the Centre for Spatial Data Infrastructures and Land Administration, Department of Infrastructure Engineering. He has worked on a wide range of research and development projects advancing the science and practice of land administration in Australia and overseas.

Abbas Rajabifard is a Professor and Director of the Centre for Spatial Data Infrastructures and Land Administration (CSDILA) at the University of Melbourne. He is also Discipline Leader of Geomatics, and Leader of the Future Infrastructure Research Program, at the Faculty of Engineering and IT. Prof. Abbass is an Advisory Board Member of the United Nations Academic Network Global Geospatial Information Management (UNGGIM).

CONTACTS

Hamid Hosseini
Department of Infrastructure Engineering, University of Melbourne, Australia
VIC 3010 AUSTRALIA
Web site: https://www.linkedin.com/in/hamid-hosseini-01aa57153/

Behnam Atazadeh
Department of Infrastructure Engineering, University of Melbourne, Australia
VIC 3010 AUSTRALIA
Web site: https://findanexpert.unimelb.edu.au/profile/653223-behnam-atazadeh

Abbas Rajabifard
Department of Infrastructure Engineering, University of Melbourne, Australia
VIC 3010 AUSTRALIA