An Approach To Improving Traditional Cadastral Data Management Based On Deep Learning And A Temporal GIS Model

Cadastral data management is an important part of land administration.In terms of management approach,it can be divided into traditional cadastral management and digital cadastral management.Traditional cadastral management is generally tailored with a static paper document,scanned map or handwritten records,which is difficult to apply automatic queries and tracking changes of a single parcel.Digital cadastral management using information technology,with single parcels as management units,can be more efficient for data management and query.It is certainly that digital cadastral management is the mainstream of cadastral management approach in the future.Based on a complete literature review and investigation,there are still some problems in progress,and two of them are studied in detail in this research.First,from a global perspective,a large number of countries or regions such as Tanzania,Kenya,Uganda and Zambia are still using traditional cadastral management approaches.If turning to digital cadastral management,one of the challenges is the lack of effective methods for digitising historical and present cadastral data stored in paper format,which makes it difficult to turn a large number of static maps into digitalised objects with parcel as a spatial unit.Second,the existing digital cadastral management puts more emphasis on storage,query and visualization of present data,such as three-dimensional-real-scene cadastral data management systems.However,it is insufficient for tracking historical changes of parcel properties(such as spatial range,ownership,land use type,etc.),which limits the application of digital cadastral data management.In view of the above,this dissertation research has proposed an approach involving a deep-learning-based method for digitising paper-based cadastral data and a temporal GIS model with full consideration of temporal characteristics of parcel property changes for improving cadastral data management in systems.The proposed approach is implemented to examine its validity and performance using real sets of data from Tanzania used as the case study.The proposed digitisation method consists of four automated sections.The first section cleans cadastral plans’ areas which do not contain parcels and extracts administrational information such as plan numbers and names of the cadastral blocks.The second section uses deep neural networks of L-CNN and ResNet-50 to detect cadastral parcels’ line segments and their labels from cadastral plans,respectively.Training and validation of L-CNN to attain its objective used 968 samples,and ResNet-50 used 106,000 samples.The third section combines the results of L-CNN and applies a topological check and corrections for parcels whose lines are not well connected.The fourth section removes redundant vertices of the digitised parcels and uses the joined and disjoined concepts to populate their labels.Furthermore,it uses nesting techniques to update numbers of the cadastral plans and names of cadastral blocks into parcels.The new temporal GIS model,extended from the Land Administration Domain(LADM)model,uses a funnel data structure.Three-time constructs of the decision,valid and transaction times are proposed with the designed model to manage cadastral parcels and their underlying information taken during parcellation processes.Changes in parcels are identified by(1)versioning previous parcels and(2)adding new parcels whose parents are not available at the base state.The storage of data is designed in one system with three inter-dependent sub-databases.The first sub-database stores active cadastral parcels that also exist in the real world and its management is done using the Base State with Amendment(BSA)model’s approach.The second sub-database stores inactive(parent)cadastral parcels that existed in the real world in the past times and their relationships with child parcels.The parent-child relationships of parcels are created using the Space-TimeComposite(STC)model’s approach.The last sub-database stores all non-spatial properties of the cadastral parcels,such as their land use and owners’ information.All three datasets and their time constructs are organised to interact in systems using geo-relational database constructs.The performance of the proposed approach is examined using 493 cadastral plans with a total of 55,172 parcels.For the digitisation method,the correct results of parcel boundaries and labels by the L-CNN and ResNet-50 are 94.45%and 75.6%,respectively.About 2000 parcels of the above are combined with their land use noted from cadastral plans and landowners downloaded from Mockaroo,as required with a temporal GIS model of the approach proposed in this study.All datasets are then transacted in a PostgreSQL database system with their changes simulated at four different times.The designed database system is then interacted with four spatiotemporal queries whose solutions cannot be fully executed with the existing data models.The first query wanted to understand details of all new parcels transacted in a period,such as six months.The second query wanted to verify the order of changes applied in one parcel,which was not developed for a while.The third query wanted to prove the ability of the designed model in restoring any previous state of the cadastral parcels and their key properties of land use and owners’ information.The last query wanted to track parent parcels of more than one generation and their land use and owners’information.All exemplified queries are successfully executed.The successful results of digitising paper-based cadastral data and retrieving temporal information of the cadastral data as stored with a temporal GIS model confirm that the proposed approach could be implemented and help to improve cadastral data management in systems.The main contributions of research of this dissertation are based on its new approach proposed to improve cadastral data management,especially for countries that are still using paper-based cadastral systems.1.It proposes a new automatic method to speed up digitising paper-based cadastral data by involving deep learning techniques.This method,in particular,is developed to detect and digitise cadastral parcels and their labels from the cadastral plans using deep learning.Its use can be extended in many other related tasks in the entire field of GIS.2.It proposes a new temporal GIS model with multiple time constructs and semantical abstractions.It is an extension of the exiting models.With this model,existing cadastral data management systems can efficiently and accurately support more queries on historical changes and properties of parcels.