University of Twente Student Theses


Visible cadastral boundary extraction using VHR remote sensing images: A deep learning approach

Tareke, Bedru Wudye (2022) Visible cadastral boundary extraction using VHR remote sensing images: A deep learning approach.

[img] PDF
Abstract:The United Nations sustainable development goal (SDG), particularly indicator 1.4.2, promotes tenure security for all, special attention given to the poor and vulnerable groups. Despite this, majority of the global population lacks access to formal land registration systems. It is exacerbated by the expensive and time consuming nature of conventional systems of land registration such use of high precision ground survey methods. Due to its flexibility and affordability, several countries have recently adopted fit-for-purpose land administration (FFP-LA) approach. FFP-LA promotes use of country wide aerial orthophoto, incremental improvement of measurement accuracy and using general boundary principle to speed up the process of land rights mapping. In recent years, the feasibility of using image-based mapping and automatic feature extraction techniques for (semi) automated cadastral boundary extraction has been actively investigated as an alternative to the traditional mapping methods. Many GIS applications, including cadastral mapping, require vector representation of spatial objects. Although deep learning networks such as FCN have achieved remarkable result in segmenting remote sensing images, the final map of the segmentation output is in raster format. Inspired by FFP-LA, this research has investigated the usability of deep learning methods to extract cadastral boundaries from very high resolution (VHR) remote sensing images directly in a vector polygon format. The proposed method utilizes FCN based network specifically UNET as a backbone to perform the segmentation task on the input image followed by the frame field method which simultaneously learns directional information of each pixel on the image. Through multi-task learning, frame field output improves the quality of deep segmentation model while providing structural information to facilitate the polygonization process. The polygonization method inspired by active contour model (ACM) takes the segmentation mask and frame field information as input and iteratively optimizes edges and corners to align seamlessly with the reference data. The experiments were conducted in the rural part of Ethiopia where agriculture is the predominant land use. The proposed method was evaluated on two different orthophotos based on images taken from UAV and aircraft platforms with 0.11 and 0.29 cm. resolution, respectively. The results are evaluated and reported using PoLiS and IoU metric. Polygons predicted on UAV orthophotos have higher similarity with the reference polygon than polygons predicted on aerial orthophotos with PoLiS distance of 2.81 and 8.64, respectively. Furthermore, a higher mean IoU of 0.84 was achieved on polygons predicted using UAV orthophotos compared to predictions on aerial orthophotos which is 0.79. Frame field based model predictions delivered simplified polygons with regular edge and corner compared to standard segmentation model. In addition, the model is tested in a different geographical area using aerial orthophoto to test its transferability. mIoU of 0.67 and PoLiS distance of 7.11 is achieved which is slightly lower than the quality of polygons predicted using UAV orthophoto. The accuracy of the model was influenced by many factors such as high crop variability inside one field, existence of invisible boundaries, confusing features such as water ways and terraces that crosses the parcel, date of image acquisition and the quality of reference cadastral data. In conclusion, the proposed method shows the possibility of utilizing deep learning methods for extracting cadastral boundaries from VHR remote sensing images in a vector polygon format that can be directly used in mapping applications with little post-processing.
Item Type:Essay (Master)
Faculty:ITC: Faculty of Geo-information Science and Earth Observation
Subject:43 environmental science, 54 computer science, 74 (human) geography, cartography, town and country planning, demography
Programme:Geoinformation Science and Earth Observation MSc (75014)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page