Introduction

1Population growth leads to an ever-increasing urban settlement characterized by buildings, roads, railways, car parks and other man-made structure surfaces such as parks or gardens (Qiu et al., 2020). This extension induces the consumption of agricultural land, the reduction and fragmentation of biodiversity and an increase in urban pollution (Béchet et al., 2018). Urbanization processes (urban sprawl, densification, etc.) also have many impacts on the quality of the soils and the gradual removal of vegetation.

2Over the past twenty years, the study of this urban sprawl has been a centre of attraction for most of the researchers in remote sensing and many of them have tried well towards extraction of impervious areas or built-up surfaces using various datasets and techniques (Xu et al., 2018). Therefore, some remote sensing products related to built-up or impervious areas have been available for several years.

3More precisely, four products providing relevant information on urban settlements can be considered comparable because of their raster format with a spatial resolution of approximately 10 to 30 m and their cover at global scales (European or global), even if they have different foci. The first product, available at a global scale, is the global human settlement layer (GHSL), which is a built-up grid at 38 m spatial resolution derived from Landsat imagery by the JRC (Joint Research Centre). Products are available for 1975, 1990, 2000, and 2014 based on supervised classification (Pesaresi et al., 2016). Since 2016, the JRC has been proposing the same map developed from Sentinel-1 images from the Copernicus program available at 20 m spatial resolution (Corbane et al., 2017). The second product is the global urban footprint (GUF), which covers the whole globe with a map of 12 m spatial resolution, where built-up area corresponds to all pixels with a high texture value derived on TerraSAR-X as well as TanDEM-X synthetic aperture radar (SAR) images (Esch et al., 2012). A classification is performed based on the speckle divergence to highlight areas subject to backscattering (Esch et al., 2013). The third product is the high-resolution layer imperviousness, which is a raster layer indicating built-up areas with a spatial resolution of 20 m, created from Copernicus high-resolution remote sensing images (Spot 5 and Indian satellite imagery). This operational product covers European countries and is produced using a per-pixel supervised classification and NDVI-based calibration completed with visual improvement. This product is updated every 5 years based on the Corine land cover database for sampling and validation with the last product date from 2018. The fourth product, covering the French national scale, is the occupation des sols opérationnel (OSO) product, which is a raster map of land cover classes produced yearly since 2016 as part of the THEIA Data and Service Centre. It is obtained by ioata2, a processing chain based on a supervised and multitemporal time series of Sentinel 2 images developed by CESBIO (the Centre for Space Studies and BIOsphere). To train their model, applied by climatic region, OpenStreetMap/Corine Land Cover/RPG (vector data available in France for agricultural areas) data are used (Inglada., 2017). Built-up areas are classified into four thematic classes at 10 and 20 m spatial resolutions (dense urban, diffuse urban, and industrial and commercial areas and roads). Additionally, ESA World Cover is a new land use/land cover product over the entire world. An analysis of this product (Venter et al., 2022) has shown that urban area precision remains low and is not adapted for precise USM mapping. In fact, urban gardens and trees are not incorporated into the “Built Area”, which causes several misclassifications and salt-and-pepper effects. Recently, Microsoft developed a building footprint product (https://github.com/​Microsoft/​USBuildingFootprints) to detect the extent of buildings worldwide.

4Detecting the urban settlement extent in rural areas where built-up areas are very sparse is irrelevant for most of these products except for GUF, which outperforms the others (Marconcini et al., 2019). Most products map urban settlements with a salt-and-pepper effect, and they cover large scales with similar pixel spacing. Moreover, although most of these products allow a frequent update with open-access image data, GUF is less frequently produced, as it was derived from the relatively expensive high-resolution TerraSAR-X and TanDEM-X SAR images. The World Settlement Footprint (Marconcini, 2020) maps settlement at 10 m spatial resolution using multimodal classification (Landsat-8 and Sentinel-1 for the 2015 product and Sentinel-2 and Sentinel-1 for the 2019 product) and decision fusion on both classifications. This product also encounters misclassifications in rural areas and large-scale networks.

5These products could be developed through the availability of many Earth observation datasets. In addition to data storage, calling and applying algorithms to these datasets must be possible. Over the past 20 years, parallel computing has been the most well-known technique for storing and exploring petabytes of data (Dean & Ghemawat, 2008; Ghemawat, Gobioff, & Leung, 2003). Moreover, these techniques allow the development and deployment of automated processing chains to offer multiple possibilities to users, such as large calculation tasks or large amounts of data processing at the same time.

6A frequently updated information as an Urban Settlement Map (USM) is thus necessary to help end-users to manage, plan and/or develop cities in this context.

7With this amount of data (images and value-added products), on-demand web service or (semi)automatic processing chains implemented on HPC centres must be developed to easily execute remote sensing algorithms and reduce the downloading of data on personal computers.

8In this context and to propose a relevant map with the urban settlement extension as a morphological area reducing the salt-and-pepper effect due to per-pixel processing, the URBA-OPT processing chain has been developed. With our processing chain, we aim to identify more precisely urban settlement, including built-up areas with roads, railway, car parks. This processing chain is adapted for rapid calculation on high-performance computing clusters and uses object-based and machine learning methods based on massive streams of Sentinel-2 imagery. The USM is then calculated in the national territory between 2018 and 2021 and is ongoing for 2022. The product named URBA-OPT, which maps urban settlements, will be downloaded on the web portal A2S as a product of the THEIA Data and Service Centre, a part of the national e-Infrastructure Data Terra.

9This paper describing the processing chain and detailing the results for one year is divided into five sections. Section 2 focuses on the study site with a selection of cities and datasets. Section 3 describes the five steps of the URBA-OPT processing chain. In Section 4, we present the results only for one year and discuss them to end with a conclusion and several perspectives in Section 5.

Study site and datasets

Study area

10The experiments and results in this paper focus on metropolitan France (including Corsica), which covers 535000 km². Its climatic areas (oceanic, continental, Mediterranean and mountainous) largely explain why France's landscape is so diverse: from coastal plains in the north and west to the mountain ranges of the Alps in the southeast, as well as the Massif Central in the south and the Pyrenees in the southwest. Additionally, the islands of Corsica lie off the Mediterranean coast (Fort and André, 2013; Terasmaa et al., 2019; Yang et al., 2020).

11To illustrate the results, seven major cities, with a population between 300,000 and 2,200,000 inhabitants, spread over all of France have been chosen: Nantes, Strasbourg, Montpellier, Toulouse, Lyon, Marseille, and Paris (Figure 1).

Datasets and the A2S high-computing infrastructure

12The processing chain uses the high-resolution layer imperviousness (HRL Imperviousness) for training data (and for first cross-validation). This USM map is open-source data developed by Copernicus and available throughout Europe, including Turkey. It comprises a grid with a 20 m spatial resolution and aims to map the degree of imperviousness of each pixel from 0% to 100%. Currently, HRL Imperviousness has been produced for 2006, 2009, 2012, 2015 and 2018. Several preliminary tests using the Urban Atlas or Corine Land Cover database compared to the HRL product have shown the relevance of the preprocessed HRL Imperviousness in both classes (impervious/not impervious) to training an RF model to extract urban areas. On the basis of this finding, we chose to consider the binarized HRL product as adapted training data to extract at 10 m spatial resolution ‘urban settlement’, defined as built-up areas but also network infrastructure and artificialized surfaces (car parks, gravel, etc.). HRL has been binarized according to a threshold of 0 to get every urban settlement. This correspond to our definition and the objective of this study and the methodology will be discussed in the next section.

13The Sentinel-2 mission is composed of two satellites, Sentinel-2A and 2B (Table 1), launched in June 2015 and March 2017, respectively. They have a high revisit frequency of five days over the equator and two-thirds of a day near mid-latitudes, which is important for mapping land cover dynamics. Each sensor has 13 spectral bands with different wavelengths, from the visible to the shortwave infrared, at different spatial resolutions. The satellite data used in this research are level 1C and are downloaded via different download hubs to have constant access to all available data in all areas of France.

14Since 2018, URBA-OPT has been implemented in the A2S 'Application Satellite Survey' scientific resource platform integrated into the high-performance computing (HPC) servers of the University of Strasbourg to offer fast on-demand or on-stream processing. A2S is a high-computing infrastructure of THEIA-land (Baghdadi et al., 2015).

15To gain some time during the classification process, we also selected several thematic masks to hide seas and oceans.

URBA-OPT: a processing chain for massive streams of S2 images

16The methodology part involves the overall workflow (Figure 2) of the proposed URBA-OPT processing chain operated by A2S infrastructure. First, Sentinel-2 Level 1C data covering France are collected. Second, reference data are preprocessed for each Sentinel-2 tile to build training and validation datasets. Third, a segmentation step allowing the grouping of pixels into homogeneous segments is applied, and the urban settlement areas are extracted based on random forest, a supervised learning method. The outputs of classification are described, and a final product is calculated. Finally, several metrics are calculated for global-scale evaluation and each tile classification.

17This chain has been developed using only open sources Python libraries, such as OrfeoToolBox (OTB) and GDAL, which have been recoded to be scaled to HPC.

Sentinel-2 image and reference data preprocessing

18This step mainly comprises two operations: downloading and computing masks for each image according to three requirements and ground reference tiling. During the first step, the processing chain requests four servers to find and download the Sentinel-2 image needed: Scihub (Copernicus), creoDIAS (Copernicus), Google, and PEPS (Plateforme d'Exploitation des Produits Sentinel). Three requirements must be satisfied for URBA-OPT to download images: (1) cloud cover under 10%, (2) 99% or more data cover over the image (i.e., corresponding to the swath of the satellite), and (3) an acquisition range (e.g., January 2019 to December 2019). Users can also specify product numbers if they know which image they want to process. The chain downloads every image that meets all requirements. We applied the Sen2Cor processing chain to correct the images from the atmospheric disturbances.

19Four spectral bands were used to construct the machine learning algorithm, band 2, band 3, band 4, and band 8 (Table 1), and two spectral indices were calculated: the normalized difference vegetation index (NDVI) and the brightness index (BI). These indexes are widely used when mapping land cover and USM using machine learning approach (Jin et al., 2018). The NDVI is built from the red (R) and near-infrared (NIR) channels. The normalized vegetation index highlights the difference between the visible red band and the near-infrared band. NDVI values range from -1 to +1, with negative values corresponding to surfaces other than vegetation cover, such as snow, water, or clouds, for which the reflectance is higher in the red than in the near-infrared.

20The BI is calculated from the red (R) and near-infrared (PIR) channels and has values between 1 and 100. The BI represents the average brightness of the image. This index is therefore sensitive to the brightness of the floor, which is related to built-up areas and bare soils. It is most characteristic of the albedo. This index makes it possible to identify urbanized surfaces that contain sparse vegetation and are therefore very reflective.

21The ground reference (HRL IMP) was then clipped using a tile extent and binarized for each pixel value strictly above 0% imperviousness. In fact, through our definition, we consider every value strictly above 0% of imperviousness as a USM. HRL thresholds can be summarized as follow:

• 0% Imperviousness: This value represents completely natural areas with no artificial surfaces. Examples include dense forests, natural grasslands, and wetlands where there is no human-induced cover.

• 25% Imperviousness: This represents areas where a quarter of the surface is covered by impervious materials. Examples include suburban residential neighborhoods with gardens, driveways, and small roads.

• 50% Imperviousness: This level indicates areas where half of the surface is impervious. Examples include urban residential areas with a mix of houses, roads, and commercial buildings, as well as small parking lots.

• 75% Imperviousness: This value corresponds to highly developed areas. Examples include dense urban centers with significant road networks, large buildings, and extensive infrastructure.

• 100% Imperviousness: This represents fully built-up areas with no permeable surfaces. Examples include large industrial complexes, commercial centers, and major urban infrastructure like airports or large-scale commercial zones.

22We thus obtained a binarized product composed of two unique values: zero corresponding to the non-waterproofed or non-urban surface, and one corresponding to urban settlements and artificialized areas to keep all the imperviousness values to respect the USM definition.

23Because of the restrictive parameters used to select Sentinel-2 images, we chose not to implement cloud or shadow masks. In addition, on the basis of several tests, we detected that FMASK has several false positives when masking an image, especially when objects are very bright (e.g., industrial areas and roads). The contribution of multitemporal imagery allows us to eliminate the errors of the unmasked pixels.

Supervised and object-based classification

24A supervised object-based image analysis (OBIA) process is adopted with a first step of segmentation followed by a random forest learning algorithm to classify images. First, the dataset with the selected spectral bands and indices is segmented using a mean shift algorithm. This algorithm extracts the maximum from a density function (Fukunaga et al., 1975) and has been optimized for high computing. After several tests, we used a spatial and a range radius of 20 and a minimum size of segments of 200. these values are consistent with the subject of study, which is USMs from Sentinel-2 images.

25The image containing the final segmentation is converted to a shapefile. A class, urban or non-urban, is assigned to each segment in this file according to a threshold that defines the proportion of HRL Imperviousness binarized (explained in the section below) in a segment. After several tests, this value was set to 40%; thus, each segment containing 40% or more of its total HRL Imperviousness binarized is classified as urban (Figure 3). We made several tests to determine the optimal thresholds for classifying urban and non-urban segments (Puissant et al., 2019). These tests revealed that thresholds significantly higher than 40% result in an excessive number of false positives, failing to detect many genuine urban areas, while thresholds significantly lower than 10% result in an excessive number of false negatives. This technique enables us to consider the entire USM, including specialized but vegetative areas such as parks and sports complexes. On the same principle, a segment containing a maximum of 10% of its ground reference area is set as non-urban. The dataset is then randomly divided to obtain 40% of the total segments in the file for training and validation. In total, 24% of the segments are used for training and 16% for validation (Puissant et al., 2019). These values remain theoretical and empirical, as some tiles can contain only a few urban areas (e.g., coastal tiles). Thus, we decided to add a parameter to set a minimum percentage of urban samples in the dataset. We decided to set it empirically at 30%, which is a good compromise to have enough training samples, even for coastal tiles.

26In the second step, a random forest learning algorithm, which is one of the most robust and widely used algorithms in remote sensing processing chains, is used (Belgiu et al., 2016). It reduces overfitting (Inglada et al., 2017) and significantly increases the computational speed compared to other approaches, such as the support vector machine (SVM). Random forest (Breiman, 2001) is based on the concept of a learning strategy. It is an ensemble classifier that combines the predictions of several base estimators to obtain the final prediction. Random forest is composed of multiple decision trees that split the data into smaller groups to point under one label. This process is repeated on each smaller group until each node contains similar samples. During the decision phase, each decision tree votes on the data to set a label on the prediction sample. These algorithms improve generalizability and robustness, which is important for remote sensing image analysis. The literature explained that random forest is not very sensitive to the choice of parameters (Pelletier et al., 2016), so we decided to implement our algorithm with 1000 decision trees. We calculated the mean, median and standard deviation over each bands and indexes before applying the random forest classifier.

27This methodology is applied to each selected Sentinel-2 tile. The selections depend on the parameters defined above for each year. Then, for each ‘millesim’, we obtain several mono-date classifications for each tile.

Output classification, final product

28On the output of the classifier, a probability map and a binarized product are produced after the rasterization of the output polygons file from the random forest to 10 meters spatial resolution. A pixel tending towards 1 is more likely to belong to the USM class than a pixel approaching -1. Binarization is performed on the probability map using a threshold of 0.4 to be more restrictive than binarization at 0. This threshold was defined after several tests on single-date classifications. The user can also change this parameter in the processing chain.

29To propose an annual product and to reduce the uncertainty of the results, a fusion method at the decision level (Zhang, 2010) is applied by a majority vote between all binary classification results for each tile. A majority vote consists of choosing the most represented value per pixel to eliminate noise or classification errors (Jeon and Landgrebe, 1999).

30Finally, mosaicking is performed, and then the product is cut according to the French border to obtain the final product for a year.

Quantitative and qualitative evaluation

31To quantitatively assess the quality of the result (binary map), we propose to use a confusion matrix (1) and a statistical file (2) containing a set of related metrics (Maxwell et al., 2021), such as USM omission or USM commission error.

Producer’s accuracy (PA)

32The producer’s accuracy (also known as recall) is a statistical indicator represented in each cell of the confusion matrices. It indicates the fraction of well-ranked pixels relative to the ground reference. It can be calculated by dividing true positive values by the sum of true positives and false negatives.

User accuracy (UA)

33The user accuracy (also known as precision) informs about the fraction of well-classified pixels in the classified image. It can be calculated by dividing true positive values by the sum of true positives and false positives.

F-Score

34The F-Score (or F-measure) represents the harmonic mean between AU and PA. It is calculated by dividing twice the product of PA and UA by the sum of PA and UA.

35To complete the evaluation step and as an appendix to the results of the URBA-OPT processing chain for France, we compared our results with three other products described in Part 1: global human settlement layer (GHSL), global urban footprint (GUF) and occupation des sols (OSO). The fourth urban class available in OSO was grouped to obtain a binary product for comparison with URBA-OPT. A simple addition of two products allows us to propose a new map where the pixels can take four values:

• 0: unclassified pixel

• 1: common pixel for URBA-OPT and the chosen ground reference

• 2: pixel only present on the ground reference

• 3: pixel only present on URBA-OPT

36Through these new maps, the surfaces classified by one product or another product can be mapped and thus the difference between each product can be evaluated. Without a ground reference for the national territory, these maps allow a qualitative evaluation of the underestimation or overestimation of urban surfaces between products. In addition, metrics will be calculated between URBA-OPT and the three other urban products.

37To complete this assessment step, we also decided to randomly select 300 pixels in the urban class and compare it with a visual ground control based on Sentinel-2 (Esch et al., 2013). In order to make it more relevant, this selection is performed through three geographic zones (Figure 4), following a fusion of French regions. The southern zone mainly covers the country's mountainous areas, and the eastern and the western zone with respectively a coastal and a continental influence.

Results

38In this section, the results for the classification of USMs in France with the URBA-OPT processing chain are detailed in two subsections. First, they are presented at a global scale by proposing a map of France and a quantitative analysis by Sentinel-2 tile. Then, they are compared with the three other urban products resulting from the state of the art for seven French cities.

Global analysis

39For each year (between 2018 and 2021), a global map with USMs in red is produced, for instance, in 2018 (Figure 5). For the evaluation at a global scale, confusion matrices are calculated for all Sentinel-2 tiles in France. Three metrics are thus displayed, the recall, the precision and the F-Score described in Part 3. Figure 6 shows a map of France for each of the metrics along with a histogram of the values. Note that these values are calculated using HRL Imperviousness as the reference product.

40The recall values are mostly between 80% and 100%. The accuracy is lower, ranging from 75% to 85%, and the F-Score has values close to 85%. On the other hand, two tiles in the south of Corsica have accuracy and F-Score values between 0% and 20%. These values can be explained by the low presence of USMs on these two tiles, which biases the machine learning model. In fact, finding training segments for USM is complicated by the fact that the tiles only partially intersect the island. The lowest values occur in rural and mountainous areas in central, western, and eastern France. These maps are very similar for each year and are not presented in this paper.

Local analysis

41The results of comparisons with the three other state-of-the-art products are shown in Figures 7 and 8 for seven cities (Paris, Marseille, Lyon, Toulouse, Nantes, Montpellier and Strasbourg). A focus was placed on peri-urban spaces with strong heterogeneity in the built environment to be the most representative in USM detection.

42As shown in Figures 7 and 8, for all cities, compared to URBA-OPT, the GHSL and GUF products tend to underestimate highly reflective surfaces corresponding to activity areas. For instance, in the northeast of the Paris subset (centred on Charles de Gaulle Airport) or in the Montpellier subset (with a circle), a large part of the runways and tarmac is not extracted. This underestimation is also largely observed in the Toulouse and Lyon subsets, where the suburban areas are not well detected, especially for the GHSL product.

43Another observation is that, for most cities where natural and agricultural areas are often merged with urban surfaces, the OSO product overestimates urban surfaces. This result is specifically visible in the Marseille subset, where the mineral and vegetation areas to the northeast are overestimated for OSO, while they are not extracted on the other products. Figure 8, which shows the calculated extracted surfaces, confirms the global tendencies of GHSL and GUF to underestimate reflective surfaces such as large infrastructure or activities and OSO to overestimate urban surfaces in the periphery of cities where vegetation is more abundant.

44Tables 2, 3 and 4 show metrics calculated by considering OSO, GHSL and GUF, respectively, as ‘reference’ products to complete the analysis of results. URBA-OPT is then statistically compared to each product. The metrics are calculated and displayed for both classes, USM (1) and non-USM (0).

45We notice that compared to OSO, the UA values for USMs are close to 1 for all test sites, and the PA values are low and tend towards 0. Thus, the vast majority of USM surfaces classified by URBA-OPT are also detected by OSO.

46This analysis is the opposite for GHSL and GUF products (Tables 3 and 4). We note low UA values for all the test sites, while PA values are higher and close to 1. Thus, most USMs detected on GHSL and GUF are also detected on URBA-OPT.

47These statistical results show an underestimation of USM surfaces for GUF and GHSL products, which is confirmed by the visual results (Figures 7 and 8). Indeed, the low UA scores coupled with the high PA scores indicate an underestimation of the extracted surfaces for the reference product. Conversely, high UA scores coupled with low PA scores indicate an overestimation of the studied surfaces for the reference product. The latter observation is true for OSO, where the UA values are high while the PA values are low. This finding is also seen in the visual results, where a high commission of the USM surfaces is visible. OSO confused bare soil with USMs at most of the study sites.

Table 2: Comparison between URBA-OPT and OSO for each city presented

City	User Accuracy		Producer Accuracy		F-Score
	0	1	0	1	0	1
Paris	0.85	0.97	0.99	0.70	0.95	0.81
Marseille	0.76	0.92	0.95	0.68	0.85	0.78
Lyon	0.89	0.93	0.96	0.83	0.92	0.88
Toulouse	0.84	0.93	0.97	0.69	0.90	0.79
Nantes	0.73	0.97	0.96	0.75	0.83	0.85
Montpellier	0.76	0.89	0.87	0.79	0.81	0.84
Strasbourg	0.90	0.90	0.98	0.64	0.94	0.75

Table 3: Comparison between URBA-OPT and GHSL for each city presented

City	User Accuracy		Producer Accuracy		F-Score
	0	1	0	1	0	1
Paris	0.95	0.68	0.90	0.83	0.92	0.75
Marseille	0.91	0.81	0.90	0.83	0.91	0.82
Lyon	0.96	0.59	0.79	0.90	0.87	0.71
Toulouse	0.99	0.24	0.78	0.90	0.87	0.38
Nantes	0.84	0.89	0.90	0.83	0.87	0.86
Montpellier	0.92	0.60	0.69	0.89	0.79	0.72
Strasbourg	0.95	0.71	0.94	0.75	0.95	0.73

Table 4: Comparison between URBA-OPT and GUF for each city presented

City	User Accuracy		Producer Accuracy		F-Score
	0	1	0	1	0	1
Paris	0.98	0.61	0.88	0.91	0.93	0.73
Marseille	0.94	0.74	0.87	0.86	0.90	0.80
Lyon	0.98	0.70	0.84	0.94	0.90	0.80
Toulouse	0.98	0.60	0.87	0.92	0.92	0.73
Nantes	0.89	0.86	0.88	0.87	0.89	0.87
Montpellier	0.87	0.85	0.85	0.87	0.86	0.86
Strasbourg	0.97	0.74	0.95	0.82	0.96	0.78

48Figure 9 shows a graphic with the surface areas calculated for each subset of cities. Note that OSO classifies more USMs than the other three products. We also notice that GHSL and GUF detect as many USMs as OSO and URBA-OPT for the highly urbanized study sites (Nantes, Montpellier), but they detect only a few residential areas containing a high density of green spaces (Toulouse, Lyon).

49This analysis confirms the previous results by showing that URBA-OPT and OSO detect more USM surfaces than GHSL and GUF. This finding is especially true for the three southern cities (Marseille, Toulouse, and Montpellier), where the detected surfaces are almost twice as important for URBA-OPT and OSO than for GHSL and GUF. This result can be explained by how the data are processed. Indeed, GHSL is produced at 38 m spatial resolution and then resampled at 10 m to be compared to our product. The quality of the result is thus strongly degraded because its resolution is reduced by almost a factor of 4. GUF is produced at 12 m spatial resolution, which is similar to our product URBA-OPT. Additionally, radar imagery is used for its production, which explains the several omissions, especially within the USM patch.

50To complete the analysis of qualitative and quantitative results, we also compared the results of URBA-OPT outside cities.

51We observe overestimation specifically in mountain areas (Figure 10) where surfaces are close to mineral and highly reflective surfaces, very similar to urban surfaces. In addition, the slope effect also makes detection difficult, as the reflectance corresponds to the mean of multiple surfaces but also the effects of altitude and sun-facing conditions are important. In contrast, most of the urban areas are correctly detected with a very low omission. These misclassifications are also identified on the other state-of-the-art products. In addition, at a spatial resolution of 10 m, USM omissions are rarely observed when settlement is very sparse and not visible at 10 m (Figure 11).

Ground Campaign: Random points analysis

52After the ground campaign, the results showed good USM classification for the eastern and western parts of France, with accuracy scores above 0.90 (Table 5). On the other hand, we noted several classification errors for the southern part of the country (Table 5), which could be explained by the strong presence of mountainous landscapes, whether in the Pyrenees, the Alps or the Massif Central. These results confirm the initial visual analysis (Figures 10 and 11) carried out on these same landscapes. The slope effect and the high reflection of bare soil and rock surfaces in the mountains (error of commission) make the extraction of USM a complex task. Also, some misclassification occurred in rural areas, especially in the southern France where dense trees in urban parks are mission (error of omission). For the two other zones, we reported misclassification in the border of cities, where pixels are often a mix of several surfaces (e.g. bare soil, impervious surfaces, trees …).

Table 5: Results for ground campaign for the three geographical areas compared to URBA-OPT product.

		Reference (visual interpretation)
	USM accuracy (in %)	Eastern France	Western France	Southern France
Classification	Eastern France	91	/	/
	Western France	/	94	/
	Southern France	/	/	86

Conclusion and perspectives

53This paper presented a processing chain named URBA-OPT to extract an urban settlement from a massive stream of Sentinel-2 imagery. This method, using open-source libraries and optimized for rapid calculation on high-performance computing clusters, is fully automatic with some parameters defined by users. This method is based on a supervised machine learning algorithm (RF) in an object-oriented approach that allows the production of a global urban settlement map with some metrics on each tile for evaluation.

54The originality of the method lies in the fully automatic processing of a massive stream of Sentinel-2 imagery and the object-oriented supervised approach using an existing raster to train the classification model. For European cities, the HRL Imperviousness product of Copernicus is used. With the Sentinel-2 images as an input in a processing chain, the objective was to develop a method adaptable to other cities with higher dynamic urbanization. The proposed processing chain can then be used with a training dataset coming from another urban product and, for instance, GHSL or WUF available at a world scale. The processing chain can then be launched on-demand on a selection of Sentinel-2 tiles on a country or only on an S2-tile for city analysis.

55The results at the French scale showed that the USM produced by URBA-OPT provides a more accurate delineation of urban settlement than the other state-of-the-art products, such as GHSL and GUF, available at the European scale, and OSO, available only in France. Qualitative and quantitative results showed that this product offers a good compromise between the over- and underestimation of urban surfaces even if some omission errors are noted, especially in mountainous or very sparse areas. With production each year and easy access to the THEIA Land and Services Data Centre, comparing ‘millesims’ can be very useful and relevant for detecting changes in urban sprawl. However, this product is not recommended for calculating surfaces (in ha, for instance) because of the misclassification detailed in the results. Indeed, to quantitatively assess the increasing urban surfaces, products or topographic databases with a higher spatial resolution could be more adapted.

56Our annual USM serves as a tool for providing a broad overview of urbanization trends, offering insights for urban planning and environmental management. By highlighting general patterns of urban sprawl, stakeholders can make informed decisions regarding infrastructure development, land use planning, and environmental conservation. While our product may not capture every fine-grained change in urban areas, it effectively illustrates the larger-scale expansion and contraction of USM. This perspective is particularly useful for policymakers and urban planners who need to understand long-term trends in urban growth and its implications for sustainable development.

57For annual and regular production (one per year), an identified problem with the current method is the update of the training dataset dating to 2018. Until 2021, and as the urban dynamics are not higher in France, evaluating the other millesim was very similar at a global scale. Then, a first approach to reducing the bias could be to use the result of the USM at time N-1 for learning the RF model. In fact, urban landscapes change due to various factors, necessitating regular updates to training data. While a classifier trained on a specific year's data can be reused, integrating newer data ensures accuracy and reflects current urban configurations, making the model more robust. Another approach could be to combine all the state-of-the-art products and keep only the common surfaces for the learning step. Other works have sought to improve the extraction of urban settlements by using multimodal imagery by combining Sentinel-2 and Sentinel-1 or by using other classification models based on a deep learning approach, which gives interesting initial results for urban fabric classification (Wenger et al., 2022).

Data availability statement

58Sentinel-2 L1C data have been downloaded through several portals: Scihub (Copernicus), creoDIAS (Copernicus), Google, and PEPS (Plateforme d'Exploitation des Produits Sentinel). Ground reference data (HRL) were downloaded through the Copernicus web interface (https://land.copernicus.eu/​pan-european/​high-resolution-layers/​imperviousness).