*1.2. Study Objective*

The main goal of our research is to develop an approach that simplifies image-based cadastral mapping to support the automated mapping of land tenure. We pursue this goal by developing an automated cadastral boundary delineation approach applicable to remote sensing imagery. In this study, we describe our approach in detail, optimize its components, apply it to more varied and larger remote sensing imagery from Kenya, Rwanda and Ethiopia, test its applicability to cadastral mapping, and assess its effectiveness compared to manual delineation.

We previously proposed a semi-automated indirect surveying approach for cadastral mapping from remote sensing data. To delineate roads from UAV imagery with that workflow, the number of clicks per 100 m compared to manual delineation was reduced by up to 86%, while obtaining a similar localization quality [21]. The workflow consists of: (i) Image segmentation to extract visible object outlines, (ii) boundary classification to predict boundary likelihoods for extracted segment lines, and (iii) interactive delineation to connect these lines based on the predicted boundary likelihood. In this study, we investigate improvements in all three steps (Figure 1). First, for step (i), we filter out

small segments to reduce over-segmentation. Second, for step (ii), we replace hand-crafted features and line classification based on Random Forest (RF) by Convolutional Neural Networks (CNNs).

**Figure 1.** Boundary Delineation workflow proposed to improve indirect surveying. This study optimizes image segmentation, questions whether Random Forest (RF) or Convolutional Neural Networks (CNNs) are better suited to derive boundary likelihoods for visible object outlines, and introduces additional functionalities for the interactive delineation.

Finally, for step (iii), we introduce more intuitive and comprehensive delineation functionalities. While we tested the previous workflow on two UAV ortho-images in Germany, and delineated road outlines, we now test our workflow on imagery covering much larger extents and compare the results to cadastral boundaries. In this study, we test our workflow on UAV and aerial imagery of 0.02–0.25 m resolution from Kenya, Rwanda and Ethiopia covering 722 visible parcels.

Our new functionalities for the interactive delineation address cases for which the boundary classification fails, or is not necessary. Boundary classification comes into play in cases of over-segmentation, when many object outlines exist. Then, the delineator has to choose which lines demarcate the cadastral boundary. Support comes from the lines' boundary likelihood predicted by RF or CNN. In this study, we introduce functionalities that allow connecting image segmentation lines to cadastral boundaries, regardless of their boundary likelihood.
