**5. Discussion**

Considering the constraints of the invisible cadastral boundaries, land administration professionals perceived that a 40 to 50 percent automatic delineation would be very significant in reducing time and labor involved in cadastral mapping practices [26]. This goal is reached by FCN with a 0.4 m tolerance in both Busogo and Muhoza, indicating good generalization and transferability of the proposed approach in cadastral boundary mapping. From Section 4, the numerical results and visual results proves and supplements each other, suggesting two main findings: (1) The true positives detected by the FCN are mainly visible boundaries like fences and strips of stones; and (2) gPb–owt–ucm and MRS have high recall, while FCN has high precision and better overall performance.

Compared with alternative edge detection and image segmentation approaches, FCN achieved better overall performance. The reason lies in the strong feature learning and abstraction ability of FCN. Lacking abstraction ability, standard edge detection and image segmentation cannot fill the semantic gap between the high-level cadastral boundary concept and low-level image features. As a result, gPb–owt–ucm and MRS achieved high recall but low precision. In other words, gPb–owt–ucm cannot determine cadastral boundaries from all the detected contours, while MRS cannot eliminate over-segmentation caused by the spectral differences within one cadastral parcel. It is also worth noticing that FCN performs supervised boundary detection, while MRS and gPb–owt–ucm are both unsupervised techniques. This may explain their differences in the abstraction ability. Being supervised, the proposed FCN-based detector is trained to detect cadastral boundaries and to disregard other irrelevant edges like building outlines.

FCN can supply high precision, while gPb–owt–ucm and MRS can supply high recall. Therefore, for further study, we can consider a combination of these methods. We can combine them in two ways. The first one is to involve the output of gPb or MRS along with UAV images as input for FCN. FCN has a strong feature learning ability. It is possible that FCN can determine cadastral boundaries from the outputs of gPb or MRS, hence increasing both precision and recall. The second way is to apply an

approach similar to [27], where the boundary map of FCN and the probability map of gPb are linearly combined and followed by the owt–ucm procedure to extract connected boundaries.

## **6. Conclusions**

The deep FCN proposed in this research is capable of extracting visible cadastral boundaries from raw UAV images. Experiments carried out on both study sites achieved an F-score around 0.5. Very clean and clear boundaries were extracted by the proposed method, avoiding the effect of messy building contours. In both study sites, the proposed method performed better than contending algorithms. The knowledge of the local experts is needed to correct the extracted boundaries and include them in a final cadastral system. We conclude that the proposed automated method followed by experts' final correction and verification can reduce the processing time and labor force of the current cadastral mapping and data updating practices.

So far, the proposed technique is mainly suitable when a large proportion of boundaries are visible. Detecting invisible boundaries, i.e., not demarcated by physical objects, from remotely sensed images is obviously extremely challenging. In other researches, invisible boundaries are identified by manual digitization via post-processing steps. The its4land project proposed a QGIS plugin (https://its4land.com/automate-it-wp5/) which supports an interactive, semi-automatic delineation to expedite the process. In the beginning of this research, the authors also attempted to fill this research gap by considering the fact that a cadastral boundary often lies in between two buildings. We tried to introduce building information as an additional input to train the FCNs for identifying the invisible boundaries. However, our experimental results obtained so far showed that there was no obvious improvement by adding building information in cadastral boundary detection. Further research is required in this direction. In future research, we will consider to improve the capability to detect cadastral boundaries (visible and invisible) using Generative Adversarial Networks (GANs) [28]. Within this framework, the generative model and discriminative model form an adversarial training, which is sharpening the training process by focusing on the most critical samples to learn. This approach is expected to improve the accuracy in boundary detection tasks which are often characterized by scarce training data.

**Author Contributions:** M.K., C.P., and X.X. conceptualized the aim of this research; X.X. wrote the majority of the paper; C.P. and M.K revised and edited the paper over several rounds; and X.X. set up and performed the experimental analysis using parts of the software codes developed by C.P. and under the supervision of both C.P. and M.K.

**Funding:** This research received no external funding.

**Acknowledgments:** The authors would like to thank the team of "its4land" research project number 687828 which is part of the Horizon 2020 program of the European Union, for providing UAV data for this research.

**Conflicts of Interest:** The authors declare no conflict of interest.

## **References**


© 2019 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).
