*4.1. Working Hypothesis: Improving Boundary Mapping Approach*

Compared to our previous workflow [21], we improved each of the three workflow steps. For image segmentation, we remove the previous need to reduce the image resolution for images larger than 1000 × 1000 pixels, and we introduce a filtering step that allows us to limit over-segmentation by reducing the number of segment lines by 80%. For boundary classification, we implement Convolutional Neural Networks (CNNs) and thereby eliminate the previous need for Random Forest (RF) hand-crafted feature generation. For interactive delineation, we develop two additional delineation functionalities ('Connect around selection', 'Connect lines' endpoints'), we develop an attribute selection for the least-cost-path functionality ('Connect along optimal path') and redesign the GUI to be more intuitive. While we previously tested our approach on road outlines only, we now show advantages compared to manual delineation for cadastral mapping, which includes various object types. The number of clicks per 100 m compared to manual delineation was previously reduced by 76% and 86%, respectively, when delineating roads from two UAV images. Now we applied our approach to delineate 378 visible cadastral boundaries from UAV and aerial imagery of larger extents, while requiring on average 80% fewer clicks compared to manual delineation.
