*4.2. Working Hypothesis: CNN vs. RF*

By reformulating our problem to be solvable by a CNN, we have investigated integrating a more state-of-the-art approach in our previously proposed boundary delineation workflow [21]. A deep learning CNN was assumed to be superior to a machine learning RF, as CNNs require no hand-crafted features, and can be trained incrementally. This starting hypothesis holds true: Even though pre-trained on images from computer vision, transfer-learning a CNN on remote sensing data provided more accurate predictions for boundary likelihoods compared to RF. Our successful integration reduces the effect of possibly meaningless or biased hand-crafted features, and increases the degree of automation of our approach. However, when conducting the final workflow step, i.e., interactive delineation, we found that we seldom made use of the boundary likelihood. We reduced over-segmentation, due to post-processing the image segmentation. This, in combination with new interactive delineation functionalities, is more effective than manual delineation for regular-shaped parcels surrounded by visible boundaries. The delineation functionality that uses boundary likelihood is beneficial for long or curved boundaries, which was rare in our study areas.
