Sentinel-2 Data for Land Cover/Use Mapping: A Review

Abstract

The advancement in satellite remote sensing technology has revolutionised the approaches to monitoring the Earth’s surface. The development of the Copernicus Programme by the European Space Agency (ESA) and the European Union (EU) has contributed to the effective monitoring of the Earth’s surface by producing the Sentinel-2 multispectral products. Sentinel-2 satellites are the second constellation of the ESA Sentinel missions and carry onboard multispectral scanners. The primary objective of the Sentinel-2 mission is to provide high resolution satellite data for land cover/use monitoring, climate change and disaster monitoring, as well as complementing the other satellite missions such as Landsat. Since the launch of Sentinel-2 multispectral instruments in 2015, there have been many studies on land cover/use classification which use Sentinel-2 images. However, no review studies have been dedicated to the application of ESA Sentinel-2 land cover/use monitoring. Therefore, this review focuses on two aspects: (1) assessing the contribution of ESA Sentinel-2 to land cover/use classification, and (2) exploring the performance of Sentinel-2 data in different applications (e.g., forest, urban area and natural hazard monitoring). The present review shows that Sentinel-2 has a positive impact on land cover/use monitoring, specifically in monitoring of crop, forests, urban areas, and water resources. The contemporary high adoption and application of Sentinel-2 can be attributed to the higher spatial resolution (10 m) than other medium spatial resolution images, the high temporal resolution of 5 days and the availability of the red-edge bands with multiple applications. The ability to integrate Sentinel-2 data with other remotely sensed data, as part of data analysis, improves the overall accuracy (OA) when working with Sentinel-2 images. The free access policy drives the increasing use of Sentinel-2 data, especially in developing countries where financial resources for the acquisition of remotely sensed data are limited. The literature also shows that the use of Sentinel-2 data produces high accuracies (>80%) with machine-learning classifiers such as support vector machine (SVM) and Random forest (RF). However, other classifiers such as maximum likelihood analysis are also common. Although Sentinel-2 offers many opportunities for land cover/use classification, there are challenges which include mismatching with Landsat OLI-8 data, a lack of thermal bands, and the differences in spatial resolution among the bands of Sentinel-2. Sentinel-2 data show promise and have the potential to contribute significantly towards land cover/use monitoring.

1. Introduction

The global land cover is rapidly changing due to anthropogenic activities (e.g., agricultural expansion and urbanisation) and natural processes (e.g., flooding) [1,2,3]. These changes impact human life, and hence effective monitoring mechanisms are needed for the sustainable management and utilisation of natural resources (e.g., forests, water). The development of satellite remote sensing technology has revolutionised the approaches in monitoring the natural and human resources on the Earth’s surface, and this technology makes it possible to monitor large areas [4]. Since the launch of the first satellite, which was dedicated to monitoring the surface of the Earth (Landsat 1) on 23 July 1972 [5], the scientific community has seen several satellites with both commercial (e.g., IKONOS, SPOT) and non-commercial (e.g., Landsat, Sentinel) business models. These satellites produce different remotely sensed data for different applications, such as forest, urban, natural hazard and agricultural monitoring. The available remotely sensed data, based on a free access policy (e.g., Landsat), have been playing an important role in monitoring natural resources and various ecosystem processes, such as forest dynamics, especially in developing countries where financial resources for the acquisition of remotely sensed data are limited [6,7].

In 2014, the Copernicus Programme, which is under the European Space Agency (ESA), launched the first Sentinel satellite—Sentinel-1A. So far, the Copernicus Programme has launched several satellite missions including Sentinels-1, 2, 3 and 5. One significant contribution of the Copernicus Programme was the launch of the multispectral instruments—Sentinel-2 satellites. The Sentinel-2 constellation is made of twin satellites; Sentinel-2A and Sentinel-2B (https://sentinel.esa.int/web/sentinel/missions/sentinel-2). After launching Sentinel-2A on 23 June 2015, the first images were received a few days later [8,9]. Sentinel-2B was then launched on 7 March 2017. Sentinel-2 satellites carry onboard multispectral imaging instruments (MSI) with the capabilities of recording 13 wide-swaths bands [9]. The primary objective of the Sentinel-2 mission is to provide high-resolution satellite data for land cover/use monitoring, climate change and disaster monitoring [9,10]. The other important objective of Sentinel-2 is to complement the other global satellite programmes such as the Landsat and SPOT (Satellite Pour l’Observation de la Terre) satellite programmes by ensuring continuity in monitoring the dynamics on Earth’s surface [8,11,12,13].

The scientific community, government agencies and private sectors have used Sentinel-2 data for different applications, such as agricultural, urban development and forest monitoring [12,14,15]. For example, Bruzzone, et al. [16] cited land cover/use monitoring as one of the essential applications for Sentinel-2 data. Other examples of the important application of Sentinel-2 include the development of a high spatial resolution (20 m) map for Africa for 2016 (i.e., CCI Land Cover—S2 prototype Land Cover 20m map of Africa) [17], the Copernicus Land cover services high spatial resolution maps (https://land.copernicus.eu/pan-european/high-resolution-layers), and the new pan-European high spatial resolution land cover/use maps (http://s2glc.cbk.waw.pl/) [18]. Countrywide high spatial resolution maps based on Sentinel-2 data have also been produced for Germany, Belgium, Bulgaria, Belgium, and Greece [19,20].

Since the beginning of the 15th century [21], the Earth’s surface has experienced rapid changes, which are driven by agricultural expansion [22], climate change [23] and rapid urbanisation [24]. These changes need monitoring instruments such as Sentinel-2 remotely sensed data to assess the status of the Earth’s surface continuously and to inform decision-makers about future changes. Moreover, long-term (>5 years) land use/cover change monitoring by Sentinel-2 has the potential of strengthening existing policies by providing accurate and timely information [8,25]. For example, Sentinel-2 data is playing an important role in monitoring the progress of achieving the Sustainable Development Goals (SDGs) [26]. Like Landsat images [27,28], Sentinel data from all missions can be accessed free of charge on Copernicus Open Access Hub (https://scihub.copernicus.eu/). Hence this data has the potential to contribute to land cover/use monitoring in many parts of the world, especially in countries where financial resources for acquiring remotely sensed data are limited [29].

There have been many studies based on Sentinel-2 data since the launch of these satellites in 2015 [30,31,32]. However, to the best of our knowledge, there has not been a review study dedicated to the importance of Sentinel-2 land cover/use monitoring, highlighting its uses and effectiveness. Therefore, the objectives of this review are to, (1) assess the contribution of Sentinel-2 data in land cover/use monitoring, and (2) explore the utilisation and opportunities of Sentinel-2 images. At the end of this review study, the best practices for using Sentinel-2 data are recommended. This review will be useful to new users of Sentinel-2 images, especially that Sentinel-2 data is relatively new (i.e., five years) as compared to other free access images such as Landsat (i.e., over four decades of operation).

2. Methods for Searching Literature

A systematic approach to database search was used to explore literature in three databases; Google Scholar, Scopus and ScienceDirect by using approaches suggested by Blaschke [33] and Ma, et al. [34]. The literature search focused on articles on Sentinel-2 and land cover mapping. The search terms were combined using the Boolean operation (OR, AND) to search for specific literature relating to Sentinel-2. The initial search was done using the terms “Sentinel-2” AND “land cover” or “Sentinel-2” AND “landcover.” A further search was done on the specific applications of Sentinel-2 images using search terms such as “Sentinel-2” AND “Forest,” “Sentinel-2” AND “Agriculture”, and “Sentinel-2” AND “urban” (

Table 1. The criteria used for the systematic literature search.

Criteria	Details
Keywords	“Sentinel-2” AND “land cover”, “Sentinel-2” AND “landcover”, “Sentinel-2” AND “Forest”, “Sentinel-2” AND “Agriculture”, and “Sentinel-2” AND “urban”
Document type	Journal articles, book chapters and conference proceedings, reports
Language	English
Publication period	2015 to 2020

). Only the literature published between 2015 and 2020 was considered because Sentinel-2 was launched in 2015. Other records were also identified through other means, such as recommendations by experts, and these records mainly included reports on the current applications of Sentinel-2.

The literature search was refined by removing the double records from the three databases. The articles which were not having the search terms in the title, keywords and abstract were also removed from the list. The literature considered in this study, includes published articles, conference proceedings and book chapters. Due to the many important applications taking place, which have not been published in peer-reviewed journals, some reports were also considered. Although other articles did not meet this criterion, they were also consulted, especially on the background and characteristics of the Copernicus Programme and Sentinel-2 images.

3. Results

The initial literature search retained 4990 articles. The first refinement which considered articles which had the search terms in the title, abstract and keywords returned 1154 articles. To be concise, and reduce the number of articles under focus, a total of 204 articles remained after considering articles which directly address the topic of Sentinel-2 land cover/use mapping. Since the search terms had duplications (due to different databases searched), 36 articles were corrected for double records and then 168 articles were considered in the final analysis. With the addition of 9 other records from other sources, the total articles considered was 177 (Figure 1).

3.1. Characteristics of the Reviewed Studies

About 60% of the literature accessed were from journals such as Remote Sensing of the Environment, Applied Earth Observation and Geoinformation, Remote Sensing, and Photogrammetry and Remote Sensing. Other studies (40%) where from journals which do not directly deal with remote sensing (e.g., Applied Geography, Forest Ecology and Management) and other sources including conference proceedings and technical reports. Conference papers were an important component of this study considering that Sentinel-2 is a newer satellite programme compared to other programmes such as Landsat; therefore, there is a lot of debate presented in scientific conferences.

3.2. Trends of Published Articles on Sentinel-2

Figure 2 compared the trends for Sentinel-2 and Landsat-8, a satellite sensor that was launched 2 years earlier than the launch of Sentinel-2. The general search without refinement showed an upward trend for both Sentinel-2 and Landsat-8. Landsat had consistently higher numbers of published articles; however, the trends were similar for the two sensors. Blaschke [33] indicated that it is common to expect the number of published materials to increase for new sensors or contemporary processing methods due to increasing usage.

The studies considered in this review were conducted in many countries across the world including countries from Asia, America, Europe and Africa (Figure 3). The distribution of the articles shows that most of the articles on Sentinel-2 were done in Europe, specifically in countries such as German, Romania, France, Bulgaria and Turkey. These studies focused on different topics on Sentinel-2 land cover classification which ranged from pre-processing to practical applications (e.g., forest and urban area monitoring).

4. Discussion

4.1. Background of ESA Copernicus Sentinel Programme

The ESA Sentinel missions are coordinated by the Copernicus Programme which is under the European Union Earth’s observation programme [13]. All the operations of ESA are funded by the European Commission in partnership with ESA, EU member states and EU Agencies. In 1998, the ESA and EU introduced the Global Monitoring for Environment and Security (GMES), which was later called the Copernicus Programme in 2014 [13]. The ESA and the EU have established a funding programme for the Copernicus Programme to provide financial support for the period between 2014 and 2020 to manage the satellite networks and to launch new satellites [10,25]. The Copernicus Programme has three main objectives: (1) to produce and disseminate information to support EU global policies for environment and security, (2) to provide a platform for stockholders, providers and users for dialogue and collaboration, and (3) to provide a legal, financial, organisational and institutional framework for the smooth function for ESA satellite missions.

The Copernicus Programme has strategic plans for developing seven satellite missions; four of these satellites (Sentinel-1–3, 5) (It is important to note that Sentinel-4 is still under construction. The two satellites under Sentinel-4 are due to be launched in 2023 and 2030). constellation have already been launched [8,9]. The first Sentinel satellite, Sentinel 1A, was launched on 3 April 2014 and carries a C-band synthetic aperture radio detection and ranging (radar) instrument. The remotely sensed data collected by Sentinel-1 satellite has a wide range of applications which include sea and land monitoring, emergency response due to environmental disasters, and economic applications (e.g., urban expansion). Sentinel-3 is dedicated to oceanography, and the first satellite of the Sentinel-3 constellation (Sentinel 3A) was launched on 16 February 2016. In 2017, the Copernicus Programme launched Sentinel-5 to monitor air pollution.

Each Sentinel mission is based on the constellation of two satellites which reduces the revisiting time, and hence providing data in the shortest possible time [10,25]. The Sentinel programme has been implemented in three phases including the pre-operation (2008–2010), initial operation (2011–2013), and full operation (2014 and beyond) [25]. Under the strategic plans for the Copernicus Programme, other satellites will be launched starting with Sentinel-4 and will go beyond Sentinel 6 in the near future [25].

4.2. Overview of Sentinel-2 Mission

Sentinel-2 Earth observation satellites carry multispectral imaging systems and acquire optical images [8,35]. Sentinel-2 satellites are operated by ESA, and the satellites were manufactured by a consortium led by Airbus Defence and Space (Airbus DS). The mission supports several services and applications such as agricultural monitoring, disaster management and land cover/use classification [15,36,37,38].

4.2.1. Properties of Sentinel-2 Data

Sentinel-2 data has a global coverage of the Earth’s land surfaces from 56° S to 84° N, coastal waters, and the whole Mediterranean Sea [8,28]. Compared to the swath width of Landsat missions of 185 km [29], the Sentinel-2 mission has a wide swath of 290 km field of view [30]. The orbit for Sentinel-2 is sun-synchronous at an altitude of 786 km, 14.3 revolutions per day, with a 10:30 a.m. descending node at the equator [17]. This local time for the equator bypass was selected to minimise cloud cover and ensuring suitable sun illumination. The bypass time for Sentinel-2 satellite matches the Landsat’s and SPOT’s bypass time, combining the Sentinel-2 data with historical images to build long-term time-series data [8], which are necessary for natural resource monitoring.

Sentinel-2 offers improved data compared to other low to medium spatial resolution satellite images (e.g., Landsat), especially in temporal and spatial resolution [39]. The 13 bands for Sentinel-2 images have spatial resolutions ranging from 10 to 60 m (

Table 2. Characteristics of ESA Sentinel-2A and -2B satellite images [9,49].

		Sentinel-2A		Sentinel-2B
Spatial Resolution (m)	Bands	Central wavelength (nm)	Bandwidth (nm)	Central Wavelength (nm)	Bandwidth (nm)
10	Band 2—Blue	492.4	66	492.1	66
	Band 3—Green	559.8	36	559	36
	Band 4—Red	664.6	31	664.9	31
	Band 8—NIR	832.8	106	832.9	106
20	Band 6—Red edge	740.5	15	739.1	15
	Band 7—Red edge	782.8	20	779.7	20
	Band 8A—Narrow NIR	864.7	21	864	22
	Band 11—SWIR	1613.7	91	1610.4	94
	Band 12—SWIR	2202.4	175	2185.7	185
60	Band 1—Coastal aerosol	442.7	21	442.2	21
	Band 9—Water vapour	945.1	20	943.2	21
	Band 10—SWIR—Cirrus	1373.5	31	1376.9	30

) [8,32]. The visible and the near-infrared (NIR) bands have a spatial resolution of 10 m, the infrared bands have 20 m spatial resolution and the other bands have 60 m (Table 1). The 10 m spatial resolution makes Sentinel data to have the potential for detailed exploration of the Earth’s surface (e.g., urban sprawls and agriculture). The other valuable characteristic of Sentinel-2 data is its high temporal resolution of 5 days [40]. This temporal resolution improved from 10 to 5 days after the launch of the second twin satellite, Sentinel-2B, which makes the two satellites to operate at 180° orbit phase [8,9].

Due to the high temporal resolution (i.e., 5 days), land cover/use changes that take place within a short period (e.g., fire incidences, floods, volcanic eruptions) can be monitored effectively. For example, Phiri, et al. [41] used Sentinel-2 images to monitor floods in the Beira region of Mozambique, while Verhegghen, et al. [42] monitored fire burnt areas using Sentinel-2 images in the Congo Basin. The application of Sentinel-2 data to monitor these incidences, which happen over a short period, makes the images more useful in countries where floods (e.g., Malawi, Mozambique and Zimbabwe), cyclones and fire incidences are common [41,42,43]. Furthermore, other programmes, such as the UN-Spider initiative have helped estimate the extent of flooding on a large-scale using Sentinel-2 products [44].

The Sentinel-2 sensor has low radiometric calibration uncertainty that makes the radiance of the images to produce reliable results. Gorroño, et al. [45] and Gorroño, et al. [46] reported radiometric uncertainty ranging from 0.03 to 0.4% for Sentinel-2 images. These values are comparable to other sensors such as the Landsat-8 [45] and thus, Sentinel-2 images have the potential to produce highly accurate information to support different applications.

The other characteristics of Sentinel-2 data that make the monitoring of the Earth’s surface more effective include the wide swath and the free access data policy. The wide swath of 290 km makes the processing of large areas much easier and more accurate with less need for data normalisation and merging [29]. Sentinel-2 data is free, making it easy for resource-constrained researchers to use the data and complement it with other free access data such as Landsat [6,30,32,47]. With many developing countries (e.g., African countries) having challenges with financial resources to secure commercial-based remotely sensed images, Sentinel-2 offers a good alternative for high spatial resolution images [6]. Sentinel-2 high spatial resolution images have already contributed to the 20 m land cover maps for Africa [17] and other regional land cover maps based on Sentinel-2 images, especially that most of the regional land cover/use maps have a spatial resolution of 30 m [1,48].

4.2.2. Sentinel-2 Data Products

Sentinel-2 data is available in different processed forms [9,25,50]. This is because Sentinel-2 MSI products undergo different stages of processing to reach a level that can be accessed by the users. The main stages include Level-0, Level-1A, Level-1B, Level-1C and Level 2A (Figure 4). Level-0 and Level-1A are not released to users and are in the form of compressed raw image data in instrument source packet (ISP) format. Level-1B product is made of granules with 25 by 23 km long of about 27 MB. Level 1B product provides radiometrically corrected imagery with Top-Of-Atmosphere (TOA) radiance values, and the product includes the refined geometry being used to produce the user accessed Level-1C products. Level-1C is made of 100 × 100 km tiles in an orthorectified format in UTM/WGS 84 projection. Using digital elevation models, Level-1C is produced in cartographic geometry (i.e., visualisable model) [9]. Level 2A production can also be processed from level 1C products by using the Sentinel-2 Toolbox [9]. From all these data products, Level-1C (Top-of-Atmospheric reflectance) and Level-2A (Bottom-of-Atmospheric reflectance) are the most commonly used products in land cover/use mapping.

4.3. Pre-processing of Sentinel-2 Images

4.3.1. Geometric Correction

The geometric correction adjusts the position of the images in line with the ground position [9,13,28,51]. The Sentinel-2 images use a physical model for geometric correction by employing ground control points (GCPs), known geographical locations for imagery referencing [25]. This model combines position, altitude and transformation information to carry out the geometric correction. An automated correlation process between reference images and GCPs is employed for geometric correction for Sentinel-2 data into a cartographic model. The reference data used for geometric correction belongs to the worldwide geo-referenced data, based on Sentinel-2 mono-spectral images [25]. The Shuttle Radar Topographic Mission (SRTM) Digital Elevation Model (DEM) is also used to improve the geometric accuracy [9].

Although Sentinel-2 has a low geometric error, Storey, et al. [52] reported that Sentinel-2 has a geolocation error of 12.5 m which is higher that the geolocation error for Landsat-8 of 12 m [53]. Storey, et al. [52] also reported that there is a sensor-to-sensor misalignment between Sentinel-2 and Landsat-8 of 38 m. To normalise this error, the National Aeronautics and Space Administration (NASA) has developed a robust harmonisation programme (Harmonized Landsat and Sentinel-2) for the two datasets to reduce this error [54,55]. So far, no known studies have focused on assessing the consistency of Sentinel data with other satellite data, including the earlier version of Landsat images (i.e., Landsat 1–7).

The atmospheric/topographic (ATCOR) software provided by the Sentinel-2 toolbox software handles the other part of the geometric and topographic correction. Topographic correction focuses on reducing effects due to shadows and surface irregularities [56]. Generally, the ESA carries validation meetings (Validation Team Meetings) to address different application challenges and to improve the accuracy of Sentinel-2 products [57].

4.3.2. Atmospheric Correction

Pre-processing improves the quality of the images by reducing the errors associated with data acquisition. Like other spaceborne optical sensors, atmospheric, topographic, shadows, and cloud cover effects also affect Sentinel-2 [58,59]. These effects have the potential of reducing classification accuracy during a land cover/use mapping [60]. It is important to note that the classification accuracy reported in this manuscript is the overall accuracy (OA). Generally, different atmospheric correction methods have been applied to Sentinel-2 data. For example, Pflug, et al. [58] tested the performance of ATCOR on Sentinel-2 images and reported that the results were similar to those of Landsat-8 and RapidEye images.

Other studies used different methods for atmospheric corrections of Sentinel-2 including Fast Line-of-sight Atmospheric Analysis of Hypercubes (FLAASH), 6S and Dark Object Subtraction (DOS) which have the potential of reducing atmospheric effects on Sentinel-2 [61]. The unique aspect of Sentinel-2 pre-processing is the availability of the Sentinel Application Platform (SNAP) or Sen2Core for pre-processing the images provided by ESA [62,63]. This platform offers an opportunity for different pre-processing including atmospheric and topographic corrections [9,59].

Due to the importance of image correction for atmospheric effects, many studies have reported the increasing development of new algorithms for Sentinel-2 atmospheric correction. Vanhellemont [64] tested the use of dark spectrum fitting atmospheric correction on the aquatic environment, and high overall classification accuracy was obtained. In a separate study, Hagolle, et al. [65] developed multi-temporal and multispectral methods for atmospheric correction to estimate aerosol optical thickness over land. Furthermore, image correction for atmospheric effects (iCOR) has also been developed, and it is effective on Sentinel-2 [66].

4.3.3. Cloud Cover Masking

Similar to other space-based optical sensors, cloud cover is common for Sentinel-2 [67]. Many studies have developed and tested different methods of cloud masking of Sentinel-2 images, including Fmasking and fusion with thermal bands from other images [67,68]. Fmasking is the most common method used for cloud masking and has been used in many studies to improve the quality of the Sentinel-2 images. For example, Frantz, et al. [68] developed an improved version of the Fmask algorithm by using a cloud displacement Index. Although different methods for cloud masking have been tested on Sentinel-2 data, studies focusing on how these methods improve the land cover/use classification accuracy for Sentinel-2 are still lacking. Other institutions such as Vlaamse Instelling voor Technologische (VITO) (https://remotesensing.vito.be/hubspot-topics/Sentinel-2) in Belgium have focused on improving the application of Sentinel-2 by developing cloud masking tools in order to improve the accuracy of Sentinel-2 data for different application such as forest, agricultural and disaster monitoring [69,70].

4.4. Land Cover/Use Classification with Sentinel-2

Sentinel-2 satellite was launched at a time when many advanced classification methods were already developed. These methods are based on both pixel [30,71] and objects [39,72]. High computing capabilities have also contributed to the advancement in land cover/use classification using Sentinel-2. Land cover/use classification of Sentinel-2 has been dominated by machine-learning approaches (Table 2) including random forests (RF), k-nearest neighbour (KNN), support vector machine (SVM) and Bayes. RF classifier is commonly used compared to the other classifiers [7,36,73].

Sophisticated machine-learning techniques such as convolutional neural network (CNN) have also been used on Sentinel-2 images [74]. For example, Segal-Rozenhaimer, et al. [75] applied CNN on land cover classification and achieved high classification accuracy of 91%. In addition, cloud-based computing has also contributed to improved land cover/use monitoring because large dataset can be analysed at a fast rate [76,77]. For example, Hiestermann, et al. [77] employed cloud-based computing using Google Earth Engine to map crops in South Africa. RF and Maximum likelihood classifiers (MLC) have also been used to produce the high spatial resolution (20 m) land cover map for Africa based on Sentinel-2 data—the CCI Land Cover—S2 prototype [17].

4.4.1. Supervised and Unsupervised

Many studies on Sentinel-2 data have shown that a supervised classification approach is applied more than an unsupervised classification approach [31]. The major reason is that many classification algorithms have been developed based on the supervised classification approach, while the unsupervised classification employs the Iterative Self-Organizing Data Analysis Technique (ISODATA) and k-means clustering as the major classification algorithms [78,79]. Supervised classification is often applied by combining object-based image analysis (OBIA) and machine-learning classifiers. However, recent studies show that many researchers are still using pixel-based classification [71,76,80]. The application of machine-learning classifiers with OBIA has shown the potential of producing high classification accuracy. However, the pixel-based approaches are still commonly used (Table 2).

4.4.2. Pixel-Based Image Analysis

The pixel-based land cover/use classification is one of the most common classification approaches applied to Sentinel-2 [31,81,82]. The literature shows that RF is the most common classifiers used for pixel-based approach (

Table 3. Selected studies on pixel and sub-pixel-based land cover/use classification of Sentinel-2 images.

Study	Land Cover/Use	Classification Method	Classifier	Accuracy (%)
Clark [83]	Bareland, built-up area, vegetation, crops	Subpixel	MESMA, RF	74–84
Colkesen, et al. [85]	forest, soil, water, corn, barren, impervious surfaces	pixel-based	CCF	87–95
Degerickx, et al. [84]	Roof, pavement, soil, shrub, tree	Sub-pixel	MESMA	57–90
Denize, et al. [7]	Crop residues, bare soil, winter crop, grassland	Supervised pixel-based	SVM, RF	81
Forkuor, et al. [62]	Agriculture, urban	Supervised pixel-based	RF, SVM, ANN	87–92
Fragoso-Campón, et al. [86]	Forest, shrubs, water, rocks	Supervised pixel-based	RF	73–79
Gašparović, et al. [87]	Water, built-up, bare soil, forests	Supervised classification	MLC, ANN	83–91
Glinskis, et al. [88]	Oil pam	Supervised pixel-based	MLC	60–70
Immitzer, et al. [8]	Maize, onion, sunflower, sugar beet	pixel-based	RF	65–76
Khaliq, et al. [89]	Water, Cabbage, Maize, built-up	Supervised pixel-based	RF	91
Kussul, et al. [90]	Crops, bareland, water	Unsupervised pixel-based	ANN	88–94
Miranda, et al. [31]	Water, forest, urban bareland	Supervised pixel-based	MLC	100
Pesaresi, et al. [12]	Built-up area	Supervised pixel-based	SML	60
Rujoiu-Mare, et al. [81]	Forest, waterbodies, built-up	Supervised pixel-based	MLC, SVM	92–98
Sekertekin, et al. [71]	Waterbody, settlement, bareland, vegetation	Supervised pixel-based	MLC	78–85
Steinhausen, et al. [91]	Cropland, forest grassland, urban areas, water	Supervised pixel-based	RF	89–91
Thanh Noi, et al. [73]	Residential, impervious surface, agriculture, bareland, forest, water	Supervised pixel based	RF, SVM, KNN	90–95
Vuolo, et al. [40]	Carrot, Maize, potato Pumpkin	Supervised pixel-based	RF	91–95
Weinmann, et al. [92]	Forest, garden, Fields, settlements	Supervised pixel-based	SVM	72–80

Note that accuracy includes the range of producer’s, user’s and overall accuracy. In Table 3, RF refers to random forest, MLC is Maximum likelihood classifier, SVM is support vector machine, ANN refers to the Artificial Neural Network, MESMA is multiple endmember spectral mixture analysis, SML is Symbolic Machine-learning and CFF is canonical correlation forest.

). Due to the limitations of pixel-based approaches, such as salt-and-pepper effect or speckle—spectral noise, sub-pixel methods have been applied on Sentinel-2 data to improve the classification accuracy [83,84]. Spectral mixture analysis (SMA) is one of the robust methods which are used for land cover/use classification for Sentinel-2 to reduce mixed pixel effects. For example, Degerickx, et al. [84] applied Multiple Endmember Spectral Mixture Analysis (MESMA) to urban land cover/use and achieved an accuracy of 85%.

4.4.3. Object-Based Image Analysis

Due to the high spatial resolution of Sentinel-2 data, OBIA is mostly preferred (see Figure 5) for land cover/use classification [8,93]. However, there is still a considerable number of studies that have used the pixel-based classification approach (Table 3 and Figure 5). This is shown by the growing number of studies reporting that OBIA produces higher classification accuracy compared to pixel-based [87,94]. OBIA has been used for different land cover/use classification including water, agriculture, forests and urban areas (

Table 4. Selected studies on the object-based land cover/use classification of Sentinel-2 images.

Study	Land Cover/Use	Classification Method	Classifier	Accuracy (%)
Dong, et al. [98]	Cropland	OBIA-Classifier	RF	78–96
Clark [83]	Bareland, built-up area, vegetation, crops	OBIA-Classifier	RF	75–84
Csillik, et al. [99]	Wheat, maize, rice, sunflower, forest, unclassified	OBIA-Rule based	Ruleset	78–98
Delalay, et al. [100]	settlement, industry, water, forest	OBIA-classifier	RF, CT	61–95
Derksen, et al. [80]	Crops, road, orchards	OBIA-Contexture	Contextual	80–90
Gašparović, et al. [87]	Water, built-up, bare soil, forests	OBIA-Classifier	ANN	83– 91
Gašparović, et al. [87]	Water, built-up, bare soil, low vegetation, forest	OBIA-Rule based	Ruleset	84–91
Gómez, et al. [101]	Winter, wheat, Others, Built-up	OBIA-Classifier	RF	84–98
Heryadi, et al. [102]	forest, water body, urban, bare land	OBIA-classifier	k-NN	80–98
Immitzer, et al. [8]	Maize, onion, sunflower, sugar beet	OBIA-Classifier	RF	65–76
Kaplan, et al. [103]	Water, Forest, wetland, urban, green field, dry fields	OBIA-Rule based	Ruleset	89–90
Kaplan, et al. [72]	Water, Forest, wetland, urban, green field, dry fields	OBIA-Rule based	Ruleset	88–90
Kolokoussis, et al. [104]	Land, seawater, oil spill, possibly dissolved oil spill	OBIA-Rule based	Ruleset	72–91
Labib, et al. [105]	Built-up, water, vegetation, Shadow	OBIA-Rule based	Ruleset	67–71
Laurent, et al. [106]	Canopy, brown leaves, green leaves	OBIA-Classifier	Bayesian	96–98
Lu, et al. [107]	Plastic-mulched land-cover, crops	OBIA-Classifier	CT	88–90
Marangoz, et al. [108]	forest, water body, urban	OBIA-Rule based	Ruleset	80–88
Marangoz, et al. [109]	Bare land, forest, settlement, vegetation, water	OBIA-Rule based	Ruleset	66–76
Mongus, et al. [110]	Agriculture, forest, Water, grassland	OBIA-Classifier	Naïve Bayes	88–95
Novelli, et al. [39]	Greenhouse	OBIA-Classifier	RF	89–93
Phiri, et al. [41]	Water, built-up area, forests	OBIA-Classifier	RF	67–91
Popescu, et al. [111]	urban area, water, forest, agriculture	OBIA-Classifier	Latent Dirichlet Allocation (LDA)	-
Weinmann, et al. [92]	Settlement, industry, water, forest	OBIA-Classifier	RF	80–83
Xiong, et al. [76]	Cropland	OBIA-Google Earth Engine	SVM, RF	68–85
Zheng, et al. [93]	Roads, bareland, Forest	OBIA-Classifier	KNN, ANN, RF, SVM	70–90

Note that accuracy includes the range of producer’s, user’s and overall accuracy. In the table, RF refers to random forest, SVM is support vector machine, ANN refers to the Artificial Neural Network, LDA refers to Latent Dirichlet Allocation and CT is classification tree.

). The development of machine-learning approaches in land cover classification has contributed to the efficient performance of OBIA [95]. For Sentinel-2 data, many studies have focused on the application of the combination of OBIA with machine-learning classifiers on land cover/use classification, due to the high spatial resolution [39,93]. Other important issues, such as the effects of segmentation parameters, have not been fully tested on Sentinel-2 data [28,96,97].

4.4.4. Accuracy of Sentinel Land Cover/Use Mapping

High accuracies have been reported on the land cover/use classification of Sentinel-2 data (Table 2 and Figure 6), and most of the classification accuracies achieved are above 80%. Other methods, including the pixel-based approach, using the maximum likelihood classifier (MLC), have also proved to produce high classification accuracies [31,38,112]. A comparison of accuracies based on different machine-learning classifiers has shown that RF and SVM outperform the other classifiers (Figure 6).

The comparison of accuracies for different classifiers needs to be interpreted with caution because the performance of these classifies depends on several factors such as number of training samples [113,114], number of land cover classes [112], the type of terrain [115] and pre-processing techniques applied on the images [60]. Since the studies reported here were conducted under different conditions, they are not totally comparable.

4.5. Integration of Sentinel-2 with Other Remotely Sensed Data

Sentinel-2 has been successfully fused/integrated with different remotely sensed data to improve its applicability in land cover/use mapping and analysis. Fusion by different bands from Sentinel-2 has also been successful because of the different spatial resolution of the different bands. For example, Wang, et al. [116] fused the 20 m bands with the 10 m bands to improve the spatial resolution. The standard and simplest fusion method of Sentinel-2 data is the pan-sharpening approach. Gašparović, et al. [87] indicated that Brovey, Intensity Hue Saturation (IHS), and Principal Component Analysis (PCI) improve the classification accuracy of Sentinel-2 land cover/use classification. Fusion with other images like the Landsat thermal band has been implemented to reduce cloud cover on Sentinel-2 data [67,116].

Sentinel-2 data has also been integrated with other datasets such as synthetic aperture radar (SAR), Light Detection and Ranging (LiDAR) and other higher spatial resolution images, including Unmanned Aerial vehicle (UAV) images [10,87,117]. The success of Sentinel missions has also been shown by the complementary use of data from different Sentinel missions. Many studies have reported the integration of Sentinel-2 with Sentinel-1 data in different applications including urban land cover/use, wetland mapping and biomass assessment [7,10]. Sentinel-1 synthetic aperture radar (SAR) data offers several advantages when integrated with optical Sentinel-2 data. These benefits mainly include the improvement in land cover/use classification accuracy. Sentinel-2 has also been successfully used with Sentinel-3 data (i.e., topographic and surface temperature data), and the results have shown promise [95]. da Silveira, et al. [117] integrated Sentinel-2 with LiDAR data to classify seven vegetation types in northeast Brazil, and the results showed a significant improvement of accuracy from 49 to 61%. This was mainly attributed to the complementary role played by Sentinel-2 reflectance information and LiDAR metrics, especially in differentiating forest succession stages based on vertical attributes.

4.6. Opportunities and Challenges of Sentinel-2 Data

The past five years of working with Sentinel-2 data, since 2015, has shown great potential in land cover/use mapping and analysis [15,118]. The major factors driving the success of the Sentinel-2 data are the free access to the data, the high spatial resolution (10–20 m), the short revisit time, and the presence of the red-edge bands. Turner, et al. [6] indicated that free and open access data contributes to the increasing applications of remotely sensed data. This is similar to the NASA Landsat data, which has become a global tool for land cover/use mapping and analysis at the region and global scale because of the free access policy [27,28]. With the availability of the high spatial resolution, the scientific community and practitioners are looking forward to land cover/use maps based on the high spatial resolution (10 m) Sentinel-2 images at the national, regional and global scale.

The high resolution of the Sentinel-2 data offers opportunities for detailed land cover/use mapping at a fine scale. The visible and near-infrared bands have been used for land cover/use analysis, which needs high resolution, such as urban land cover/use classification, crop monitoring and plantation forest mapping. For example, Yang, et al. [119] investigated the season flooding in urban areas using Sentinel-2 data. Generally, there has been an increase in studies on crop monitoring and these studies range from yield assessment to crop identification [36,90]. The monitoring of small-scale forest plantation was also reported to be successful using Sentinel-2 in the Tanintharyi region in Myanmar [38]. Given that the high spatial resolution of 10 m is only on four bands (visible to near infra-red), this poses a challenge when integrating other bands, such as short-wave bands and the red-edge bands, which need to have the same spatial resolution. Many studies have also proposed methods of downscaling the low spatial resolution images to 10 m through image fusion [63,87].

The presence of the three red-edge bands has shown high applicability to land cover/use mapping, especially in vegetation land covers/uses [62,120,121]. The red-edge bands have been used to improve land cover/use classification [62] and specific mapping of land covers/uses, such as grassland monitoring [15] and land cover/use disturbances, such as oil spill [121]. Kussul, et al. [90] compared the classification results between Landsat OLI-8 which has no red-edge band and Sentinel-2 images. The results showed that the red-edge bands improved classification results by 4–5%.

Besides the opportunities offered by the Sentinel-2 data, there are many challenges associated with using Sentinel-2 data. The primary limitations include misalignment with other remotely sensed data (e.g., Landsat 8), the absence of the panchromatic (i.e., black and white) and thermal bands, and the variation in the spatial resolution of the bands. Storey, et al. [52] reported that the Landsat OLI-8 and Sentinel-2 images have a sensor-to-sensor misalignment of 38 m due to the errors in ground control points (GCPs) and Global Land Survey (GLS) framework. On a global scale, there were plans to correct the misalignments by 2018. However, users need to be cautious and apply manual correction to align the images. Furthermore, other automated methods have been developed to solve the problem of misalignment through different steps for image pre-processing before integrating Landsat OLI-8 and Sentinel-2 images [30]. NASA is also undertaking the harmonisation of Landsat OLI-8 and Sentinel-2 (HLS), which aims at obtaining corrected images through atmospheric correction, cloud and shadow masking, co-registration, bidirectional reflectance normalization and bandpass adjustment [54,55].

The 13 bands from Sentinel-2 do not include a panchromatic band nor a thermal band. For those applications which need a panchromatic band, such as pan-sharpening, such applications cannot be applied on the original Sentinel-2. Different methods using other independent bands such as single bands or by averaging the bands with a high spatial resolution (i.e., the four bands with 10 m spatial resolution), have shown promise in improving Sentinel-2 land cover/use monitoring [116,122]. Thermal bands are essential for various applications of cloud cover masking [67,68]. With Sentinel-2 imagery lacking a thermal band, other studies have suggested using thermal bands from the Sentinel-3 mission [123]. However, the thermal bands from Sentinel-3 have the low spatial resolution, and hence, need to be pan-sharpened [123].

4.7. Best Practices for Optimal Classification Accuracy with Sentinel-2

To produce the desired classification accuracy with Sentinel-2 data, several technical aspects need to be taken into consideration. These include pre-processing and selecting appropriate methods for classification. Pre-processing, such as atmospheric and topographic corrections, needs to be considered before the classification. Atmospheric correction is ideal for multiple images, especially for time series analysis. There are many approaches for atmospheric correction. However, the Sentinel toolbox offers atmospheric correction software dedicated to Sentinel-2 pre-processing. Topographic correction aims to normalise the effects due to surface irregularities and can also be corrected using the Sentinel-2 Toolbox.

Due to the increasing number of available methods for land cover/use classification, critical consideration needs to be put into place to choose the appropriate classification method. Many studies have shown that supervised OBIA, using a machine-learning classifier, produces the desired results [34,61,124]. However, several considerations have not been tested, especially the scale parameters for OBIA [8,28,110]. Other methods like spectral mixture analysis (SMA) have the potential [83,125], especially in reducing mixed pixel effects.

4.8. Specific Applications of Sentinel-2 in Land Cover/Land Use Monitoring

Sentinel-2 multispectral images have been used for land cover/use monitoring in different ways across the world. These various aspects include forest mapping [126,127,128], carbon assessments [129,130], urban area mapping [131,132] and natural hazards monitoring [41,133,134]. Other vital applications include agricultural [135,136] and water resources monitoring [77,137]. The studies on the implementation of Sentinel-2 in land cover/use monitoring have been reported in many countries across the world, including Europe, Asia, America, and Africa. Many studies that have been reported in developed countries, especially in Europe, integrated Sentinel-2 data with contemporary datasets such as LiDAR and UAVs datasets [86,117,138], while studies, based in developing countries, combined Sentinel-2 data with other open-source data such as Sentinel-1 SAR data and Landsat [139,140].

4.8.1. Sentinel-2 for Forest Monitoring

In the forest sector, Sentinel-2 products have been a powerful tool because of being useful to different applications including mapping of forest area [126,128], establishing boundaries of specific forest types [127,131], discrimination of forest types [132,141], and other applications such as leaf area index (LAI) analysis [142,143]. In all these applications (

Table 5. Summary of major forest application of Sentinel-2 imagery across the world.

Application	Specific Application	Country	Methods	Accuracy	Reference
Forest	Forest extent	Poland, China, Burkina Faso, South Africa, Madagascar, Zimbabwe, Bulgaria	Machine-learning, cloud computing	80–90%	Suresh, et al. [158], Wang, et al. [127], [126], Adjognon, et al. [131]; Nzimande, et al. [128]; Filchev [159]
	Forest types	Italy, Ghana, South Africa, Togo	Linear discrimination analysis, spectral indices, Machine learning	88–90%	Laurin, et al. [132]; Konko, et al. [160]; Puletti, et al. [141], Laurin, et al. [132]
	Species Identification	Germany, Italy	OBIA-RF, Stepwise regression	65–76%	Immitzer, et al. [8], Laurin, et al. [132]
	Forest productivity	Germany, South Africa, Southern Africa	Machine-learning (Random Forest), Invertible Forest Reflectance Model	90–92%	Mutowo, et al. [161]; Ramoelo, et al. [162]; Darvishzadeh, et al. [163]
	Growing stock	Norway, Greece, Italy, Finland	Fusion with UAV data, Linear regression	SE=3.4–5.8%	Puliti, et al. [146], Chrysafis, et al. [164], Mura, et al. [118]
	Forest Inventory	Finland, Norway	Fusion with UAV data, multivariable models	SE=3.4–5.8%	Puliti, et al. [146], Astola, et al. [165], [166]
	Wetland mapping	China, Canada, South Africa, Senegal, Ghana	machine-learning, Google Earth Engine, OBIA	83–90%	Yesou, et al. [167], Mahdianpari, et al. [168], Whyte, et al. [144]; Mondal, et al. [152]
	Leaf Area Index (LAI)	Finland, Germany, South Africa, Bulgaria	Red-edge band with Partial Least Squares Regression (PLSR), Spectral Indices	R2 = 91%	Clasen, et al. [143], Korhonen, et al. [11], Sibanda, et al. [142]; Dimitrov, et al. [169]
	Forest Fires/Wildfire	Madeira Island, Bulgaria, Congo DRC, Africa (continent)	Active fire products with SAR data fusion, New Algorithm	80–89%	Verhegghen, et al. [42]; Roteta, et al. [149]; Navarro, et al. [150], Nedkov [170], Filchev [159]
	Dryland mapping	German, South Africa	Sub-pixel classification, BiomeBGC Simulations	82%	Munyati [171], Dotzler, et al. [172]
	Grassland mapping	South Africa	Sparse Partial Least Squares Regression (SPLSR)	R2 = 59%	Shoko, et al. [156]
	Canopy cover	Finland, German	Generalized additive models, Spectral Unmixing and UAVs	RMSE = 0.05–0.42	Korhonen, et al. [11], Clasen, et al. [143]
	Forest succession	Brazil, Poland	SVM, RF, OBIA	90–97%	Sothe, et al. [173], Szostak, et al. [126]
	Forest Degradation	Bulgaria, Tanzania	OBIA-RF	R2 = 0.97, 95%	Hojas-Gascon, et al. [174], Nedkov [170]
	Forest healthy	Poland	Machine-learning	75–78	Hawryło, et al. [175]
	Forest phenology	German	Correlation with ground sensor	R2 = 0.99	Lange, et al. [176]
Biomass assessment	Above-ground biomass	Vietnam, Finland, South Africa, Zimbabwe, Italy	Machine-Learning, SPLSR, PARAS, Regression analysis	80–91%	Pham, et al. [154], Pandit, et al. [177]; Shoko, et al. [156], Majasalmi, et al. [155], Laurin, et al. [132]
	Below grown biomass	Turkey	Regression analysis, Supervised classification	R2 = 97%	Bulut, et al. [130]
Carbon Assessment	Carbon assessment	Czech Republic, Turkey, South Africa	Bootstrapped Random Forest, Regression analysis, multivariate regression models	R2 = 30–75%	Bulut, et al. [130], Naidoo, et al. [129], Gholizadeh, et al. [157]

), Sentinel-2 data have proved to be more useful than other low spatial resolution images (e.g., Landsat) because of the high spatial resolution (10 m) [144] and the availability of the red-edge band [145].

The application of Sentinel-2 images differ from region to region; however, the major difference is between the developing and the developed countries. In developed countries, such as Finland and Germany, Sentinel-2 data is increasingly applied for specific applications (e.g., forest inventory) requiring detailed analysis. However, most applications in developing countries focus on describing the extent of the forests (land cover/use). For example, Puliti, et al. [146] reported the use of Sentinel-2 in forest inventories in Norway. The increasing combination of UAVs [146,147] and LiDAR [86,117] datasets with Sentinel-2 has shown promise and was highly accurate in describing different forest attributes.

Invasive plant species have also been monitored using Sentinel-2 images. Kattenborn, et al. [147] reported the use of Sentinel-2 in combination with Sentinel-1 and UAVs to monitor three invasive species (Pinus radiata, Ulex europaeus and Acacia dealbata) in Chile. In Baringo County, Kenya, Ng, et al. [148], compared Sentinel-2 and Pléiades (2-m spatial resolution) data to produce a highly accurate vegetation map that would differentiate an invasive tree species (Prosopis) from native forest trees and mixed vegetation classes. This information is useful for effective forest management strategies. Although they observed that higher spatial resolution of Pléiades contributed to high accuracy, it was concluded that Pléiades is costly and the free of charge Sentinel-2 data provides a viable alternative as its increased spectral resolution compensates for the lack of spatial resolution.

Forest fire (wildfire) monitoring has been one of the crucial applications of Sentinel-2 imagery. With forest fires being common in most of the tropical regions [149], Sentinel-2 has proved to be a valuable tool due to the high temporal resolution of five days. Navarro, et al. [150] reported the application of Sentinel-2 multispectral images for post-fire monitoring using spectral indices in Madeira Island. Sentinel-2 images have also been used for mapping burned scars, fires severity and soil erosion susceptibility in southern France [151]. Continental level maps for Africa on fire have also been developed based on Sentinel-2 multispectral images [149]. The combination of multispectral Sentinel-2 and SAR Sentinel-1 improves the accuracy of fire monitoring [42,149]. Besides forest fires, Sentinel-2 has also been used to monitor the quality of foliage in national parks, such as the Kruger National Park in South Africa, especially after fire incidences.

Other sensitive areas requiring adequate monitoring, include the wetland ecosystems as they have been affected by different anthropogenic activities, such as an expansion of urban and agricultural areas [97]. Whyte, et al. [144] highlighted those wetland areas are sensitive to climate change, especially to the increasing temperature and the changing rainfall pattern. By using Sentinel-2, wetland maps have been developed for the Newfoundland (Canada) using Google Earth Engine. The combination of Sentinel-2 and Sentinel-1 to monitor wetland was tested in China, and this study achieved high accuracy (70–90%). In iSimangliso Wetland Park, South Africa, mapping of wetland areas was enhanced with the combination of Sentinel-2 and Sentinel-1 [144]. In line with wetland areas, mangrove forests are an important component of the wetland ecosystem and have also been affected by human activities [152,153]. Therefore, Sentinel-2 data have the potential to enhance effective monitoring of these mangrove ecosystems. For example, Mondal, et al. [152] mapped the mangrove forests in the coastlines of Senegal and the Gambia (West Africa) with high accuracy (>80%).

Sentinel-2 data has also been a critical monitoring mechanism in climate change through biomass and carbon assessment. Most of the studies have reported using Sentinel-2 data to monitor above-ground biomass [14,154,155]. In Northern Vietnam, Pham, et al. [154] reported the application of Sentinel-2 to map above-ground biomass for mangrove forests. Shoko, et al. [156] also characterised the above-ground biomass for C3 and C4 grass in Drankensberg region in South Africa using Sentinel-2 data. Although most of the studies on biomass focus on above-ground biomass, Bulut, et al. [130] determined the total biomass by considering both above- and below-ground biomass in Turkey. Many studies that focus on determining carbon quantities have used regression analysis (e.g., multivariant regression) to relate the spectral properties of Sentinel-2 and carbon quantities [129,130,157].

4.8.2. Sentinel-2 for Agricultural Monitoring

Sentinel-2 has become an important tool for monitoring agricultural activities (

Table 6. Summary of agricultural application of Sentinel-2 imagery globally.

Application	Country	Methods	Accuracy	Reference
Crop diseases	China, South Africa, Zimbabwe	Random Forest	77–94%	Zheng, et al. [189], Dhau, et al. [186]; Chemura, et al. [191]
Crop residue	Spain, Malawi	Maximum likelihood, OBIA, regression	90–97%	Andersson, et al. [82], Estrada, et al. [138], Zheng, et al. [196]
Crop type detection	Ukraine, France, Austria, Zimbabwe, Ethiopia, Bulgaria	Deep learning, Random Forest, Support vector regression	77–96%	Kussul, et al. [184], Chemura, et al. [187]; Vogels, et al. [188], Vuolo, et al. [40], Veloso, et al. [197]; Dimitrov, et al. [198]
Crop yield focusing/productivity	Belgium, Saudi Arabia, Ukraine, Zimbabwe, Mali	Deep learning, Random Forest, Maximum likelihood, Spectral indices	35–96%	Lambert, et al. [182]; Hiestermann, et al. [77], Al-Gaadi, et al. [135], Delloye, et al. [199]
Cropland extent	United Kingdom, Madagascar, Ukraine Global Dataset (Burkina Faso, South Africa, Morocco, Madagascar)	Cloud-based computing, Machine-learning, OBIA	64.4–96%	Zhang, et al. [200], Bontemps, et al. [178]; Inglada, et al. [201], Lebourgeois, et al. [136], Kussul, et al. [184]
Irrigation crop	Ethiopia, Global Dataset (Burkina Faso, South Africa, Morocco, Madagascar)	Object-based	94%	Vogels, et al. [137]; Vogels, et al. [188]
Nitrogen content	Belgium, Bulgaria	Multivariant regression	65–90%, RMSE = 0.25	Clevers, et al. [185], Dimitrov, et al. [169]
Real-time crop monitoring	South Africa	Cloud-based computing (Google Earth Engine)	-	Hiestermann, et al. [77]
Smallholder crop monitoring	Mali, Ethiopia	Supervised pixel-based, object-based	80–94%	Lambert, et al. [182]; Vogels, et al. [137]
Soil properties	Spain, France, USA	Multivariant analysis, Neural Network, TRApezoid Model	64–88%	Gao, et al. [139], El Hajj, et al. [202], Sadeghi, et al. [203]
Biophysical parameter estimates	France, Spain, Bulgaria	neural networks (NN), support vector regression (SVR), kernel ridge regression (KRR), and Gaussian processes regression (GPR)	RMSE = 0.1–0.2	Upreti, et al. [193]; Xie, et al. [192], Dimitrov, et al. [169]

). This is evidenced by the various studies, which have focused on developing global products to support agricultural activities [178,179]. For example, global agricultural maps produced by the Onesoil project in Belarus using Sentinel-2 products and machine-learning, provide useful information to farmers [180]. Bellemans, et al. [181] also reported on a global project (Sentinel-2 for Agriculture) involving African countries, such as Burkina Faso, South Africa, Morocco and Madagascar. The specific applications of Sentinel-2 in agriculture, include crop production monitoring [77,135,182], crop type mapping [183,184], irrigation agriculture monitoring [137], nitrogen content assessment [185] and assessment of crop health [186]. These studies range from small scale monitoring (i.e., field-level) [182,183] to continental level [76].

Due to the high demand for information for agriculture production, Sentinel-2 data has been used for real-time monitoring of agricultural activities [77]. An exciting development in the application of Sentinel-2 in agricultural applications is the increasing use of cloud computing applications (e.g., use of Google Earth Engine) [178]. It is important to note that cloud-based computing has become common in other applications such as forest and wetland monitoring. Therefore, Xiong, et al. [76] and Hiestermann, et al. [77] highlighted that cloud computing is useful for extensive area monitoring (e.g., national and continental level). In addition, machine-learning algorithms produce high accuracies in agricultural monitoring, especially in the discrimination crop type [187] and identifying the specific types of agricultural systems (e.g., irrigation farming) [188].

Sentinel-2 images are also commonly used for monitoring crop diseases. Zheng, et al. [189] used Sentinel-2 data to monitor wheat yellow rust in China. Heavy metal-induced stress on rice was also assessed using Sentinel-2 multispectral data in China. The results from these studies have high accuracies of over 85%, indicating that Sentinel-2 data has the potential for crop health monitoring. Dhau, et al. [186] assessed the abilities of Sentinel-2 multispectral images to detect maize grey leaf spot disease in Durban, South Africa. The results showed that including all the 13 bands for Sentinel-2 and using the RF classifier produced a high accuracy of 83%. Other applications of Sentinel-2 images that are related to plant health monitoring include detecting crop residues [82,190] and assessing the chlorophyll in crops [191].

An accurate estimate of biophysical parameters is important for a number of applications such as precision agriculture, crop productivity and soil hydrology [169,192]. Sentinel-2 images have been used for estimating biophysical parameters [193,194,195]. For example, Xie, et al. [192] used Sentinel-2 images to retrieve chlorophyll content, leaf area index, nitrogen content and leaf chlorophyll index. One of the advantages of Sentinel-2 images is the availability of the red-edge band, which is reliable in retrieving biophysical parameters [193,195].

4.8.3. Sentinel-2 for Urban Area Monitoring

The characteristics of Sentinel-2 are driving the increasing use of these images in monitoring urban areas (

Table 7. Summary of urban and natural hazard application of Sentinel 2 imagery.

Application	Specific Application	Country	Methods	Accuracy	Reference
Urban	Urban expansion	Brazil, China, Tanzania, Kenya	Spectral Indices, RF	75–92%	Gombe, et al. [212]; Ng, et al. [148]; Iannelli, et al. [216]; Tavares, et al. [204]
	Urban extent	China, Brazil	Fusion	83%	Iannelli, et al. [216]; Tavares, et al. [204]
	Rural-urban transition	Ghana	Principal Components Analysis (PCA)	-	Møller-Jensen [206]
	Informal settlement	South Africa	Cloud-based computing (Google Earth Engine)	-	Gibson, et al. [207]
	Urban surface water	China, Macedonia	Pixel-based/OBIA	80–92%	Yang, et al. [217], Yang, et al. [208], Sekertekin, et al. [218]
	Urban climate	France, Germany	Canonical Correlation Forests	69–75%	Qiu, et al. [219]
	Urban change	France	Convolutional Neural Networks (CNN)	60–91	Daudt, et al. [220]
	Urban ecosystem/forest/green space	Slovakia, Switzerland	SVM, Maximum likelihood	73–90%	Haas, et al. [140], Recanatesi, et al. [221]
	Urban heat island	Lebanon, France, German	MLC, Neural Network	82–84%	Kaloustian, et al. [222]; Qiu, et al. [223], Chunping, et al. [205]
Natural hazards	Floods	Spain, Mozambique	Spectral indices and OBIA	64–85%	Caballero, et al. [134], Phiri, et al. [41]
	Droughts	Germany, South Africa	Spectral Mixture Analysis, Biome-BGC Simulations	73–82%	Munyati [171], Dotzler, et al. [172]
	Earthquakes	New Zealand, France	cross-correlation	RMSE= 0.025–0.20	Kääb, et al. [213], Jelének, et al. [215], Stumpf, et al. [224]
	Volcanic eruption	Saunders Island, Germany	Correlation, visual assessment, Convolutional neural network (CNN)	RMSE = 0.03	Gray, et al. [214], Valade, et al. [225]

). The application of Sentinel-2 in urban areas include urban expansion [204], urban heat island [205], rural-urban transition [206], informal settlement [207], and urban ecosystem (e.g., urban forests/green space) [140]. In addition to these applications, Sentinel-2 images have also been used to monitor surface water in urban areas. For example, Yang, et al. [208] extracted surface water in Beijing, China using Sentinel-2. For most countries in Africa, urban expansion (e.g., slum settlements) due to population growth has been a major challenge [209]. Hence, Sentinel-2 provides a reliable tool for urban land use planning [210,211]. The choice of Sentinel-2 data was attributed to its high spatial resolution (10 m) to identify informal settlements accurately, and in most cases, the combination of Sentinel-2 and Sentinel-1 has produced highly accurate results.

Although Sentinel-2 data offers an opportunity for monitoring urban activities, few studies have been dedicated to urban monitoring using Sentinel-2. Most of the studies that employ Sentinel-2 data in urban monitoring mainly focus on urban expansion [206,212]. Therefore, there is a need for studies on different urban monitoring aspects (e.g., road networks, access to facilities, waste management) using contemporary high spatial resolution data (i.e., Sentinel-2).

4.8.4. Sentinel-2 for Natural Hazards

Globally, there are different natural hazards (Table 7), which affect both flora and fauna. Phiri, et al. [41] reported that natural hazards have negative impacts on infrastructure (i.e., built-up areas). These natural hazards include floods [134], droughts [172], earthquakes [213] and volcanic eruptions [214]. Water resources affect human life through floods, especially in coastal and riparian areas [41,134]. For example, Phiri, et al. [41] attempted to help policymakers in making informed decisions on pre- and post-management of floods by employing Sentinel-2 data in Beira, Mozambique. Studies on the drought area mainly focus on the vegetation growth in drought-prone regions. For example, Munyati [171] and Dotzler, et al. [172] mapped drought stress in deciduous forest communities in South Africa, and Germany, respectfully.

Sentinel-2 data has also been used to map the effects of more destructive natural disasters such as earthquakes [215] and volcanic eruption [214]. For example, Jelének, et al. [215] investigated the post-earthquake landslide distribution by using Sentinel-2 and Sentinel-1 data in New Zealand. In Sounders Island of the South Sandwich Islands, Gray, et al. [214] investigated the volcanic activities using Landsat-8 and Sentinel-2 data. The application of Sentinel-2 data in monitoring natural disasters is in line with one of the main aims of the Copernicus Programme of monitoring the Earth’s natural disasters [25].

5. Conclusions

This study aimed at understanding the contribution of ESA Sentinel-2 data towards land cover/use monitoring. The current research has shown that most studies reviewed indicated that Sentinel-2 data has the potential for land cover/use monitoring across the world. Many studies have also reported the superiority of Sentinel-2 over similar sensors such as Landsat-8. The application of sentinel-2 data differs from region to region, especially the type of data which is integrated with Sentinel-2. In developing countries, Sentinel-2 is integrated with cotemporally datasets, such as LiDAR and UAVs data, while free access data is combined with Sentinel-2 in developing countries. The major strength of Sentinel-2 is the high spatial resolution, high temporal resolution and the availability of the red-edge band [8,25]. Many classification methods have been applied with Sentinel-2 data including both pixel- and object-based approaches. However, the use of OBIA and machine learning classifiers (e.g., RF and SVM) has proved to have the great potential of improving land cover classification. The studies have also shown that due to the high spatial resolution, Sentinel-2 data can achieve high accuracies compared to other medium spatial resolution satellite images, such as Landsat. Like other optical satellite images, Sentinel-2 images are affected by cloud cover, and hence limiting its applicability in cloud prone areas. Since Sentinel-2 images are relatively new (approximated 5 years), many regions have not been tested in comparison with earlier versions of Landsat images.

Moving forward, Sentinel-2 offers new opportunities for both the private sector, government organisations, the scientific community and practitioners to increasing availability of regional, national, continental and global level land cover/use maps based on the high spatial resolution Sentinel-2 data. Future review studies can explore the applications of Sentinel data to specific regions of the world (e.g., Africa, Asia).