**1. Introduction**

In former industrialized regions characterized by a large number of brownfields and a high population density, such as Wallonia (the southern region of Belgium), offering new living spaces while limiting land take has become a challenge. The managemen<sup>t</sup> of vacant lands is then a key to urban planning, as monitoring abandoned sites can support policy and decision-making [1]. In Wallonia, many industrial sites were developed during three distinct periods between the end of the 18th century and the middle of the 20th century. However, since the middle of the 20th century, industrial sites have been increasingly abandoned, first due to the closure of coal mines, then of manufacturing and metallurgical industries. Moreover, a phenomenon of relentless de-urbanization has increasingly emptied the urban centers. This has led to the development of industrial and urban wastelands, which, depending on their origin, can vary in size from a few dozen square meters to a few dozen hectares (e.g., coal mines or blast furnaces), with 75% of them being less than one hectare. As the vast majority of these sites are located in urban areas, they negatively impact the urban fabric but also represent an opportunity for sustainable urban planning as they can be revalorized, with their reuse being a fundamental asset in land managemen<sup>t</sup> [2]. Therefore, the Walloon authorities have proposed a detailed definition for those sites and have catalogued them into an exhaustive inventory [3,4]. The redevelopment sites (RDSs) are thus defined as "property or group of properties that have been or are intended

**Citation:** Petit, S.; Stasolla, M.; Wyard, C.; Swinnen, G.; Neyt, X.; Hallot, E. A New Earth Observation Service Based on Sentinel-1 and Sentinel-2 Time Series for the Monitoring of Redevelopment Sites in Wallonia, Belgium. *Land* **2022**, *11*, 360. https://doi.org/10.3390/ land11030360

Academic Editor: Sara Venafra

Received: 28 January 2022 Accepted: 25 February 2022 Published: 1 March 2022

**Publisher's Note:** MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

**Copyright:** © 2022 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 (https:// creativecommons.org/licenses/by/ 4.0/).

to be used for an activity, excluding housing, and whose current state is against land managemen<sup>t</sup> best practices, or constitutes a deconstruction of the urban fabric" [5]. The RDS inventory, which enables potential investors and public authorities to find out about vacant land and its condition, currently contains more than 2200 sites and is available online [6]. Updating it is essential to keep a record of all the sites that have already been enhanced and provide reliable information to the actors consulting the database. Currently, this update is performed, on the one hand, by the visual analysis of orthophotos annually provided, as open data, over the entire Walloon territory and, on the other hand, by systematic field visits. These methodologies are time-consuming and costly. Indeed, the first solution requires several months of work for the analysis of all the RDSs included in the inventory; moreover, the results can only be provided once a year, and there is also a delay between the moment of data acquisition and their availability. As for the second solution, the systematic field visits, the analysis is spread over several years. However, the Walloon authorities estimate that less than 10% of the RDSs are likely to be redeveloped from one year to the other and show major changes (the three classes of interest for the administration are buildings, vegetation and soil). It is, therefore, necessary to find a way to reduce the time spent on the inventory update by providing operators with a list of sites presenting indications of significant changes that would enable them to concentrate their efforts on these sites. The problem of how to efficiently monitor redevelopment areas (usually called brownfield sites or more generally, vacant lands, although with a slightly different meaning than ours) has been examined in many studies that mostly focus on either their potential for policy-makers by using GIS data [7] or the detection of new vacant lands. In particular, remote sensing data have been used in several studies for the detection of new brownfields: Ref. [8] investigated the potential of IKONOS data in the object-oriented classification approach and Ref. [9] investigated IKONOS, QuickBird and hyperspectral data. In a recent study [10], the fusion of remote sensing images thermal data, GIS layers and citizen science data is proposed for the identification of urban vacant land. Remote sensing is also used, at a fine scale, for the detection and monitoring of hazardous substances and materials, as shown in [11].

Change detection is one of the major applications of satellite-based remote sensing data [12], and many different satellite-based change detection methods have been developed and used in recent decades. Among the most commonly used methods are algebra methods (e.g., Image Differencing, Ratioing or Change Vector Analysis), transformationbased methods (e.g., Principal Component Analysis), classification-based methods [13] and time series analysis. In [14], the authors provide a review of the different techniques, a guide to compare them by placing a clear separation of variables between the analysis unit and classification method and report that pixel and post-classification change methods remain the most popular choices. The review also presents some advantages and limitations of the different techniques. These limitations and how to overcome them have been widely studied and have led to more refined methodologies, e.g., super-resolution mapping and the analysis of mixed pixels for the improvement of land-cover class maps [15]. In addition, many other methods have recently been developed, notably based on artificial intelligence [16,17]. However, in [16], it was highlighted that supervised AI methods require massive training samples to obtain a robust model and that processing remote sensing big data requires a large amount of computational resources, which limits the implementation of the AI model. It is, therefore, crucial to choose the methodologies based, on the one hand, on needs such as the scale of the application and the thematic objectives and, on the other hand, on aspects such as the resolution of the available images and their ability to provide the required comparison features [14]. In the framework of this project, we opted for a time series analysis approach as, depending on the method, it offers a number of advantages, e.g., being able to detect abrupt and gradual changes (BFAST) or to capture subtle but consistent trends (LandTrendR), Continuous Change Detection and Classification (CCDC) being able to detect a variety of LULC changes continuously with high spatial and temporal accuracies [18]. However, in [18], the limitations of these methods are also presented,

e.g., time-consuming, requiring many resources, unsuitable for irregular observations, and some are unable to identify types of changes. It is, therefore, crucial that the choice of time series analysis method takes into account the objective of the research, and considers the need to find the change points as soon as possible in real-world applications and that there is a detection delay for many existing approaches [19,20].

Within this context, the European Copernicus program has opened, with the launch of Sentinel-1 and Sentinel-2 satellites, new opportunities thanks to their high spatial and temporal resolution. The Sentinel-1 mission consists of a constellation of two polar-orbiting satellites mounting a C-band synthetic aperture radar (SAR) imaging system. They offer a repeat cycle of six days and all-weather and day-and-night monitoring capabilities [21]. The two Sentinel-2 satellites A and B are characterized by a sun-synchronous orbit, phased at 180 to each other, and a repeat cycle of 5 days [22]. The temporal resolution of the Sentinel satellites ensures enough data to create time series [23–25], and their spatial resolution allows for the identification of landscape features [26] and monitoring urban areas [27], whereas the Sentinel-2 spectral resolution facilitates the thematic identification of land cover [28–30].

In addition to the use of SAR and optical data separately, the combination of SAR and optical data has been highlighted in domains such as vegetation monitoring [31] and urban mapping [32,33]. Combining the two types of data has the advantage of coupling features and thus overcoming some limitations, such as clouds, shadows and snow cover for the optical data. Regarding the Sentinel images, the combination has been investigated in various domains, such as forest disturbance [34], soil tillage [35] and urban mapping [36]. In [37], the use of Sentinel-1 data alone, Sentinel-2 data alone and their combined use for forest–agriculture mapping are compared.

The demand for automated operational services providing near-real time information for environmental monitoring has increased substantially in recent years, and several studies have investigated their feasibility and proposed possible implementations, mainly for natural events monitoring. In [20], the Thresholding Rewards and Penances TRP concept was applied for a near-real time forest disturbance alert system based on PlanetScope imagery, producing new forest change maps when a new image is made available. They proposed a robust statistical method to estimate forest clear-cuts, but the use of PlanetScope images makes the service costly as they need to acquire raw imagery. In [38], a near-real time automatic avalanche monitoring system based on Sentinel-1 data was presented, and an age tracking algorithm was developed, while, in [39], the focus was on burned forest areas using Sentinel-2 data. For mapping burned areas, the latter used a selection of spectral indices to compare the pre-fire and post-fire values. In [40], an automatic and repeatable plot-based change detection method, based on pre and post event Sentinel-1 and Sentinel-2 data, was designed and tested to map extreme storm-related damages. Most of the services are in the test or pre-operational phase and focus on localizing one type of change, with hindsight of events and/or using one type of remote sensing data being sometimes costly.

The goal of this paper is to present the methodological aspects and implementation details of SARSAR, a new Earth observation service for the monitoring of redevelopment sites in southern Belgium. For its deployment, a number of requirements made by the Walloon administration had to be met, namely: (i) the implementation of a straightforward automatic operational tool providing results on a regular basis (once every two months); (ii) the ability to detect changes in vegetation, buildings and soil, on a set of sites spread throughout the region's territory; (iii) the use of open-source data.

Differently from other methodologies and services mentioned above, the focus is, therefore, on providing a response to the administration need of monitoring RDSs on a regional scale and identifying the time and type of change at the site level using free and open-source technology. In brief, by exploiting Sentinel-1 and Sentinel-2 data, the service automatically detects and characterizes changes in user-defined sites of interest and provides a final change list that can be directly used by the Walloon authorities to prioritize their daily work and reduce the time needed for the inventory update.

To fulfil the free and open-source technology requirement, we exploited Terrascope, the Belgian contribution to the Sentinel Collaborative Ground Segment (CollGS), which provides access to pre-processed Sentinel data [41] and computer capacity for the execution of the process and its automation. The Sentinel Collaborative Ground Segments were created by ESA and its Member States to facilitate the access to the Sentinel data and the data exploitation. CollGS can be used for various applications, as shown by Ref. [42], who used Terrascope for geohazard monitoring.

To be able to provide a list of the RDSs that are likely to change, several steps were implemented. Considering the number of sites to be processed and the fact that aggregate information is needed for each RDS, we opted for an object-based approach. Moreover, since the number of training samples required to implement a solution based on AI would have been prohibitive, our final choice was a combination of unsupervised methodologies.

After data preparation, where the extraction of temporal features from the Sentinel time series was performed, two processes were run: first, the change point detection analysis based on the Pruned Exact Linear Time (PELT) [43], whose goal is to flag each site as changed/unchanged and to provide an estimate of the change date(s) [44] and then a rule-based classification based on threshold selection to characterize the types of changes.

Changepoint analysis is largely employed for the study of time series in many application domains, ye<sup>t</sup> it is still underexploited within the remote sensing community, due to the fact that high resolution images were not easily accessible until a few years ago. In regard to our service, changepoint detection was chosen because it serves a twofold purpose: it directly provides an estimate of the date of change, which alone constitutes valuable information for the administration, and allows us to restrict the time window within which the change classification should be performed. As regards threshold selection, it is a common procedure in algebra-based change detection [45]. The selection of the best threshold could be associated with a priori knowledge or derived from the histogram of the image [12]. The advantage of thresholding is that it can guarantee a robust near real-time approach based on fast and automated processing [34]. To the best of our knowledge, there have not been other attempts to use changepoint detection in combination with threshold-based classification for the characterization of changes in urban areas.

The paper is organized into five sections: The Materials section presents the study area, the Sentinel data used for this study via the Terrascope platform and the ground truth used for validation. The Methods section is divided into three parts: the first part explains the feature extraction and the creation of temporal profiles, the second part investigates the change detection method chosen and the third part presents the methodologies used for the classification of the changes. The last three sections are the presentation of the results, the discussion and the conclusions.
