**1. Introduction**

Cadasters, which record the physical location and ownership of the real properties, are the basis of land administration systems [1]. Nowadays, cadastral mapping has received considerable critical attention. An effective cadastral system formalizes private property rights, which is very important to promote agricultural productivity, secure effective land market, reduce poverty and support national development in the broadest sense [2]. However, it is estimated that over 70% of the world population does not have access to a formal cadastral system [3]. Traditional field surveying approaches to record land parcels are often claimed to be time-consuming, costly and labor intensive. Therefore, there is a clear need for innovative tools to speed up this process.

Since the availability of very high resolution (VHR) satellite or aerial images, remote sensing has been used for mapping cadastral boundaries instead of field surveying, and is advocated by fit-for-purpose (FFP) land administration [4]. In practice, cadastral boundaries are often marked by physical objects, such as roads, building walls, fences, water drainages, ditches, rivers, clusters of stones, strips of uncultivated land, etc. [1]. Such boundaries are visible in remotely sensed images and bear the potential to be automatically extracted through image analysis algorithms, hence avoiding huge fieldwork surveying tasks. According to FFP, boundaries measured through on-site cadastral surveys using total station or the Global Navigation Satellite System (GNSS) with a precise location are considered as fixed, while boundaries delineated from high-resolution imagery are

called general [4]. Although less spatially precise, general boundary approaches are much cheaper and faster than conventional cadastral surveys. Typically, high-resolution satellite images (HRSI) have been used for interpreting cadastral boundaries, but there are still obstacles such as high cost, cloudy or dated imagery [5]. Therefore, Unmanned Aerial Vehicles (UAV), renowned for low-cost and high spatial–temporal resolution, as well as being able to fly under clouds, are chosen as the data source for cadastral boundary extraction in this research.

The detection of cadastral boundaries from remotely sensed images is a difficult task. Above all, only visible cadastral boundaries coinciding with physical objects are detectable in the image. Moreover, as visible cadastral boundaries can be marked by different objects, spectral information alone is insufficient for the detection. In other words, there exists a sematic gap between the high-level boundary concept and low-level visual cues in the image. More reliable and informative features should be constructed to bridge the semantic gap, thus more advanced feature extraction techniques are needed.

State-of-the-art methods are mostly based on image segmentation and edge detection [6]. Segmentation refers to partitioning images into disjoint regions, inside which the pixels are similar to each other with regard to spectral characteristics [7]. Researchers claimed that segmentation-based approaches have two general drawbacks: Sensitive to intra-parcel variability and dependent on parameter selection. The latter often requires prior knowledge or trial and error [8]. Multi-Resolution Segmentation (MRS) is one of the most popular segmentation algorithms [9]. Classical edge detection aims to detect sharp changes in image brightness through local measurements, including first-order (e.g., Prewitt or Sobel) and second-order (e.g., Laplacian or Gaussian) derivative-based detection [10]. Derivative-based edge detection is simple but noise sensitive. Amongst others, the Canny detector is justified by many researchers as a predominant one, for its better performance and capacity to reduce noise [6]. More recently, learning-based edge detection stands out as remarkable progress, which combines multiple low-level image cues into statistical learning algorithms for edge response prediction [10]. Globalized Probability of Boundary (gPb) is considered as one of the state-of-the-art methods. It involves brightness, color and texture cues into a globalization framework using spectral clustering [11]. Both MRS and gPb are unsupervised techniques, hence the high-level cadastral boundary is still hard to distinguish from all the detected edges.

Recent studies indicate that deep learning methods such as Convolutional Neural Networks (CNNs) are highly effective for the extraction of higher-level representations needed for detection or classification from raw input [12], which brings in new opportunities in cadastral boundary detection. Traditional CNNs are usually made up of two main components, namely convolutional layers for extracting spatial-contextual features and fully connected feedforward networks for learning the classification rule [13]. CNNs follow a supervised learning algorithm. Large amounts of labeled examples are needed to train the network to minimize the cost function which measures the error between the output scores and the desired scores [14]. Fully Convolutional Networks (FCNs) are a more recent deep learning method. In an FCN architecture, the fully connected layers of traditional CNNs are replaced by transposed convolutions. This is the reason why these networks are called fully convolutional. As compared to CNNs, FCNs are able to perform pixel-wise predictions and accept arbitrary-sized input, thus largely reducing computational cost and processing time [15]. The superiority of FCNs in feature learning and computational efficiency makes them promising for the detection of visible cadastral boundaries, which provides the predominant motivation of this research. To the best of the authors' knowledge, this is the first study investigating FCNs for cadastral boundary extraction.

In the remainder of this article, we apply deep FCNs for cadastral boundary detection based on UAV images acquired over one urban and one semi-urban area in Rwanda. We compare the results of FCNs with two other state-of-the-art image segmentation and edge detection techniques, namely MRS and gPb [9,11]. The performance of these methods is evaluated using the precision-recall framework. Specifically, to provide better insights into the detection results, we provide a separate

accuracy assessment for visible and invisible cadastral boundaries, and an overall accuracy assessment for all cadastral boundaries.
