PhD Defence Sophie Crommelinck

automating image-based cadastral boundary mapping

Sophie Crommelinck is a PhD student in the department of Earth Observation Science (EOS). Her supervisor is prof.dr.ir. M.G. Vosselman from the faculty of Geo-Information Science and Earth Observation (ITC).

Recording land rights provides land owners tenure security, sustainable livelihood and increases financial opportunities. Estimates suggest that about 75% of the world population does not have access to a formal system to register and safeguard their land rights. This lack of recorded land rights increases insecure land tenure and fosters existence-threatening conflicts, particularly in developing countries. Recording land rights spatially is defined as cadastral mapping or cadastral boundary delineation. Cadastral boundaries can either be recorded on the ground with accurate measurement equipment (direct surveying) or be delineated along visible boundaries from remote sensing imagery (indirect surveying). Cadastral mapping is considered the most expensive part of a land administration system.

Recent developments in technology allow us to rethink contemporary cadastral mapping. Unmanned aerial vehicles (UAVs) known as drones are evolving as an alternative tool to acquire land tenure data. UAVs can capture geospatial data at high-resolution in a low-cost, transparent, and flexible manner. Imagery captured with UAVs is increasingly used in indirect surveying to accurately delineate visible cadastral boundaries. Many cadastral boundaries are visible, as they are demarcated by physical objects such as fences, walls, roads, buildings, or rivers. Furthermore, recent advances in automated detection and localization of objects from images offer new opportunities for indirect surveying: instead of delineating visible boundaries manually from low-resolution imagery, boundaries can be extracted automatically by applying image analysis to high-resolution UAV imagery. Such solutions offer the potential to improve current cadastral mapping procedures in terms of time, cost, and accuracy for the sake of worldwide land tenure security and sustainable land administration.

This Ph.D. research introduces an approach that simplifies image-based cadastral mapping. We develop an automated cadastral boundary delineation approach that is applicable to remote sensing data of high-resolution. The approach is designed for areas, in which boundaries are demarcated by physical objects and are thus visible. Areas of investigation are East African developing countries including Kenya, Rwanda, and Ethiopia.

In chapters 1 and 2, we review the state-of-the-art on cadastral mapping, boundary delineation, UAV photogrammetry, feature extraction, as well as their interactions. The review reveals that automating indirect surveying from UAV data is a recently emerging research field. In practice, indirect surveying from UAV or any other remotely sensed image appears to be rarely automated and relies mostly on manual delineation through on-screen delineation. We show the potential of automated UAV-based cadastral mapping. The review covers feature extraction methods that are synthesized into a hypothetical workflow consisting of image segmentation, line extraction, and contour generation.

In chapters 3, 4 and 5, the hypothetical workflow is implemented step by step by testing and adapting previously reviewed methods. Our results show that methods from computer vision are suitable to precisely extract object outlines demarcating cadastral boundaries. However, the methods are developed to work well on small images and not necessarily on UAV images of many more pixels. We adapt these computer vision methods and apply them to remote sensing data that cover large-scale imagery and 3D information. For image segmentation, we find that Globalized Probability of Boundary (gPb) contour detection extracts objects at completeness and correctness rates of up to 80% (chapter 3). For line extraction, we find that Simple Linear Iterative Clustering (SLIC) delineates the objects with the high accuracy provided by the UAV imagery at completeness rates of up to 64% (chapter 4). For contour generation, we implement machine learning through Random Forest (RF) classification to combine the results of gPb and SLIC. Further, we develop a procedure for a subsequent interactive delineation. Compared to manual delineation, the number of clicks per 100 m is reduced by up to 86%, while obtaining a similar localization quality (chapter 5).

In chapters 6 and 7, the approach is optimized and evaluated for cadastral mapping. While our previous evaluations use objects such as roads and buildings that potentially demarcate cadastral boundaries, the evaluations in chapter 6 and 7 compare results to the intended end product: cadastral boundaries. Furthermore, potential end users are involved and asked for feedback. The workflow’s complexity is reduced by replacing gPb and SLIC with another computer vision method, namely Multiscale Combinatorial Grouping (MCG) (chapter 6). The accuracy of boundary likelihoods predicted by machine learning is improved by 11%. The degree of automation is increased by replacing RF classification with deep learning Convolutional Neural Networks (CNN) (chapter 7). The final workflow consists of image segmentation, boundary classification, and interactive delineation. The workflow is tested on UAV and aerial imagery. We show that our approach is less tiring and more effective in terms of clicks and time compared to manual delineation for parcels surrounded by visible boundaries. Strongest advantages are obtained for rural scenes delineated from aerial imagery, where the delineation effort per parcel is requires 38% less time and 80% fewer clicks compared to manual delineation.

We aimed to automate the extraction of cadastral boundaries by translating the intelligence of a human delineator into a machine learning approach. So far, our workflow based on deep learning has not been shown to be ‘intelligent’ enough to replace a human delineator. Cadastral intelligence, which we define as deep learning frameworks for cadastral boundary delineation, requires further improvements before fully replacing manual delineation. However, we show cases where our approach is already superior to current manual delineation practices.

A successful application of our approach in a real-world use case requires more cycles of designing, developing, testing, and evaluating. This demands analyses beyond accuracy and efficiency that incorporate existing social, legal, and institutional frameworks, as well as further method development. One major limitation impeding method development throughout the Ph.D. research was the unavailability of UAV and cadastral data covering areas with many visible boundaries.

Overall, this manuscript illustrates how state-of-the-art knowledge from remote sensing, geo-informatics, photogrammetry, computer vision, and machine learning can be combined into an innovative cadastral mapping approach. By making use of synergies, we developed an approach that is superior to manual delineation when object outlines are continuously visible and coincide with cadastral boundaries. In future work, it can be adapted and transferred to real-world cadastral mapping use cases. The automated approach simplifies and speeds-up the delineation of objects from imagery. While the approach has been developed for cadastral mapping, it can also be used to delineate objects in other application fields, such as land use mapping, topographical mapping, road tracking, or building extraction. This Ph.D. research can be considered an innovative impulse for improving manual delineation in land administration and beyond.