In a context where the dynamic trend of urbanization is becoming more and more rapid, urban planning is a challenge as it involves limiting the amount of land being taken up while providing new living spaces. It is therefore important for local authorities to undertake effective measures in urban planning and development, in order to control the urban expansion, enhance the resilience of cities and preserve green spaces. In former industrial regions, such as Wallonia, there is a large number of brownfields, called "Redevelopment Sites" (RDS), which offer an opportunity for sustainable urban planning due to their redevelopment potential. These are mainly urban sites previously used for industrial activities and now abandoned. Currently, in Wallonia slightly more than 2200 RDS are listed in an inventory managed by the Walloon authorities, which required considerable time and resources to maintain up-to-date.
Within this perspective, the Sentinel satellites of the European Copernicus program are a real opportunity. Thanks to their high temporal and spatial resolution, their open access and the possibility of complementary use of different sensors, combined with the RDS inventory, they allow the implementation of an operational tool for RDS monitoring, in near-real time and over the long term.
What we propose in this study is an operational and automatic solution for the processing of Sentinel-1 and Sentinel-2 data where a combination of change detection and change classification methodologies is used to generate a final report that is directly usable by public authorities. The complete processing chain is implemented in Terrascope, the Belgian Copernicus Collaborative Ground Segment, which offers, via virtual machines, pre-processed Sentinel data and computational capacity. This enabled the automation of the process while processing and analyzing large volumes of data and images.
As far as the methodology is concerned, first, a suitable set of features (backscatter from the Sentinel-1 VH band and a selection of Sentinel-2 indices) is extracted from the data and used to create average temporal profiles for each polygon contained in the RDS vector file. The latter allows object based methodologies, one object being one RDS. Next, the PELT (Pruned Exact Linear Time) change detection method is applied to the Sentinel-1 VH and Sentinel2 NDWI2 features to determine if a change has occurred and estimate the date of the change. Finally, a classification of the changes, which is exploits the Sentinel-1 VH and Sentinel2 (NDVI, BI, BI2, SBI & BAI) features is performed to provide information on the type of change (vegetation, building and soil), the direction of the change (increase/decrease), if any, and the amplitude. This last part is composed of two distinct processes: (1) the "summer classification", which is best suited for detecting and classifying gradual changes, and (2) the "change point classification", which provides information on the type of change each time a change point is detected and an estimated date is indicated. In general, the multi-temporal approach allows to: (1) select and pre-process Sentinel data while removing outliers, (2) estimate the dates of changes, (3) characterize the changes according to the estimated dates, and (4) automatically provide results at regular intervals and/or on demand.
The results were validated based on 2 sets of ground truth created by visual analysis. The first one was obtained from orthophotos (25cm resolution) taken once a year between 2016 and 2018. The second one was created from Pleiades images (4-band pan-sharpened products at 0.5 m resolution) taken once a month over two years (2019 and 2020). The validation highlighted the relevance of the processing chain for the change identification, with a satisfactory accuracy. This applies both for the change detection (overall accuracy 62% to 85%), which also provides an estimate of the change date, and for the change classification (overall accuracy 69% to 90%), which indicates the type of change. For the change classification, the two processes "summer classification" (overall accuracy from 79% to 90%) and "change point classification" (overall accuracy from 69% to 85%) performed sufficiently well. The results showed the suitability of the combination of the two different methods, which allows us, on the one hand, to classify the type of change when a change point is detected via the PELT methodology and, on the other hand, to identify the gradual changes. Therefore, the processing chain presented in this project enables the creation, in near-real time and automatically at regular intervals, of a priority order list that highlights the RDS with the most changes and thus guides the work of the field operators, allowing them to focus on the sites that need their attention the most. This helps the Walloon authorities to manage the RDS inventory in a more efficient and reactive way, which is an important contribution to the improvement of urban planning and development measures.
As a conclusion, even if earth observation data present some limitations and challenges, such as the spatial resolution of Sentinel images, the project results show that the use of these types of data represents an opportunity to improve urban development policies. In particular, it shows an application that can be directly used by the authorities to monitor the evolution of sites that present a high potential for redevelopment. As a further use, the proposed processing chain could be used to monitor other types of sites in the field of urban planning, but also in agriculture, forestry or in disaster response.