Monitoring sustainable development by means of earth observation data and machine learning: a review

This paper presents and explores the different Earth Observation approaches and their contribution to the achievement of United Nations Sustainable Development Goals. A review on the Sustainable Development concept and its goals is presented followed by Earth Observation approaches relevant to this field, giving special attention to the contribution of Machine Learning methods and algorithms as well as their potential and capabilities to support the achievement of Sustainable Development Goals. Overall, it is observed that Earth Observation plays a key role in monitoring the Sustainable Development Goals given its cost-effectiveness pertaining to data acquisition on all scales and information richness. Despite the success of Machine Learning upon Earth Observation data analysis, it is observed that performance is heavily dependent on the ability to extract and synthesise characteristics from data. Hence, a deeper and effective analysis of the available data is required to identify the strongest features and, hence, the key factors pertaining to Sustainable Development. Overall, this research provides a deeper understanding on the relation between Sustainable Development, Earth Observation and Machine Learning, and how these can support the Sustainable Development of countries and the means to find their correlations. In pursuing the Sustainable Development Goals, given the relevance and growing amount of data generated through Earth Observation, it is concluded that there is an increased need for new methods and techniques strongly suggesting the use of new Machine Learning techniques.

• There is an increased need for new methods and techniques to process an ever-growing amount of Earth Observation data. • Machine Learning techniques are crucial in handling Earth Observation data given the enormous quantity of sources and formats.

Background
The concept of Sustainable Development (SD) has been developed in 1960 when it became evident that environmental problems can be caused by economic and industrial development. In 1972, a first report was published and presented at UN concerning SD. This report, named as the Meadows Report [1], was strongly criticised at that Open Access  [2]. Later in 1987, the Brundtland Report (BR) [3] defined the SD concept as development that meets the essential needs of the present without compromising the ability of future generations to meet their own essential needs. In 2000, the Millennium Development Goals (MDGs), established 8 objectives to tackle poverty and hunger, achieve gender equality and improve the health sector [4]. Until 2015 the MDGs [5] drove the progress of SD, including improvements in health and education services, reduced hunger and equity gaps, and higher levels of coverage in interventions with major investments [6,7]. However, it remained incomplete and in 2012, new objectives were established, designated as Sustainable Development Goals (SDGs) [5], defining 17 unique objectives, representing an urgent call to shift the world onto a more sustainable path [8,9]. Earth Observation (EO) plays a major role in supporting progress towards many of the SDGs [10,11]. According to the United Nations [12] it is advantageous using EO data such as the images from satellites to produce and support official statistics to complement traditional sources of socio-economic and environmental data. Satellite imagery may be perhaps the only cost-effective technology able to provide data at a global scale [13,14]. Such globally available data are determinant to understand the progress and contribution of underdeveloped countries concerning SD since they lack the resources to collect relevant information. The considerable amount of data, provided by EO sources, need to be effectively analysed and processed with appropriate methods and tools to provide robust indicators concerning SD.
The growth of Machine Learning (ML) field, which is constantly creating new opportunities for monitoring and evaluating humanitarian efforts, plays an essential part in the analysis of satellite images applied to SDGs. In fact, the majority of methods used for processing EO data are based on ML [11,15] given in one hand their ability to process enormous amounts of data and also because they possess unique characteristics pertaining to classification, modelling and forecasting.
The main purpose of this article is to explore and comprehend the relation between SD, EO and ML, to understand the relevance and role EO and ML play in attaining the SDGs. Figure 1 depicts the layout of this review as well as major aspects pertaining to the treatment of EO data related to the identification of SDGs.
This review highlights major methodologies and ML methods that have been successfully applied to EO data in pursue of SD. The structure of this paper is divided as follows: Sect. "Materials and methods" describes how the research was conducted, Sect. "Overview on sustainable development" presents the meaning of SD, its history, concepts and goals, followed by a brief explanation of the EO system and how it presently contributes to SDGs, in Sect. "Overview on earth observation for sustainable goals development". Afterwards, in Sect. "Earth observation using machine learning techniques", a review on the importance of ML for EO is presented and as well as their contribution for SDGs, highlighted by case studies of different ML categories applied to EO data. In addition, further considerations are addressed and discussed concerning SD, EO, ML, their relation and new paths and approaches to overcome limitations.

Materials and methods
A systematic search and analysis of published articles in peer-reviewed journals have been conducted using ScienceDirect and Google Scholar. The search has been performed using the following search topics: sustainable development or sustainable goals, earth observation and machine learning. To ensure the identification of

Overview on sustainable development
The environmental problems derived from the economic development became evident during the 1960s and a number of solutions were proposed [1,3,8]. The Limits to Growth, also known as the Meadows Report [1], was published by the Club of Rome, in 1972. It presented a computer model developed by MIT called World3, which allowed Meadows et al. [1] to explore the relationship between five subsystems of the world economy: population, food and industrial production, pollution and consumption of non-renewable natural resources [16]. The key finding has been that unlimited growth in the economy and population would lead to a collapse of the global system by the mid to late twenty-first century [1,[17][18][19]. Moreover, the sooner the world starts striving to change the growth trends, the better the chance of achieving sustainable ecological and economic stability [1,18,19]. Thus, the report advocated that the non-growth in developing countries is a response to environmental decline and the lack of its resources [1,20].
This premise became very popular among non-orthodox economists since it was translated as an attack to the capitalist economic system. On the other hand, it has also been criticised by the economists who affirm that for capitalism, it is crucial a development without boundaries. Due to that, in 1974, the Club of Rome issued another report in which it defended an organic growth (world division into different regions, each with a definite function within the world system) [2]. Since the publication of The Limits to Growth [1], a considerable number of concepts have been introduced and developed integrating ecological and economics concerns, not being consensual, until the publication of Brundtland Report (BR) in 1987 (further detailed in Sect. "Brundtland report [3]"). Table 1 presents some of the most important milestones of the path to SD until nowadays.

Brundtland report [3]
The Brundtland Report's (BR) concept of SD follows a generic definition of development that meets the essential needs of the present without compromising the ability of future generations to meet their own essential needs [3]; however, it included crucial features such as environmental preservation and meeting the basic human needs at a global scale. For those reasons, it was widely accepted as a reference for SD definition [20]. Even so, the ambiguity in the BR's concept of SD along with differing worldviews, ideologies, backgrounds, beliefs and interests has contributed to the proliferation of several explanatory definitions [23]. In an attempt to clarify and simplify the BR's concept, it became important to describe and explain the following key concepts: • Needs: necessary or basic needs (especially referring to developing countries' needs); • Technological Limitation: insufficient technological development; • Social Organisation Problems: originate an unequal allocation of income.
Later, the BR also clarified the meaning of technological growth, arguing that such progress cannot exceed the limited availability of resources [3,20].

Millennium development goals (MDGs)
In 2000, the Millennium Development Goals (MDGs) have started a global effort to tackle the indignity of the poverty problem. The MDGs [24] established eight objectives for: tackling poverty and hunger; primary education for all children; achieve gender equality; improve maternal and child health; prevent and combat deadly diseases; ensure environmental sustainability; and, global development.
Until 2015, the MDGs allowed progress in several important areas, such as: reducing poverty and child mortality; providing access to water and sanitation; improving maternal health and combatting several diseases such as HIV/AIDS, malaria and tuberculosis.
The most notable accomplishments were: the reduction of child mortality and the number of children out of school by more than half; more than 1 billion people left extreme poverty; and, HIV/AIDS infections have been reduced by almost 40%. The legacy and achievements of the MDGs provided valuable lessons and experience, and pave the way for new goals [8].

Sustainable development goals
The Sustainable Development Goals (SDGs) have replaced the MDGs in 2012 during the UN Conference on SD held in Rio de Janeiro. As a result of climate changes and other serious environmental problems, there was a need to enhance the environmental performance [25]. Hence, the main objective was to create new goals that would address the urgent environmental, political and economic challenges affecting the world [26].
Representing an urgent appeal to change the world's course into a more sustainable direction, the SDGs [27] represent a strong commitment to proceed the MDGs and tackle some of the world's most significant challenges [28].
The  [8]. The 2030 Agenda [5], which coincided with another historical agreement achieved at COP21 Paris Climate Conference [29], sets specific objectives and attainable targets for the reduction of carbon emissions, management of climate change and risks of natural disasters.
Overall, the SDGs are special because they address issues that affect the entire world and reaffirm the determination to eradicate poverty, improve the health system and reduce inequalities. Better yet, they involve all nations in building a more sustainable, safer, more prosperous planet for humanity [8,28]. To monitor and achieve the SDGs, EO became a vital part since it provides numerous benefits [10,30,31], namely: Data at

Overview on earth observation for sustainable goals development
Earth Observation (EO) covers different approaches, including the use of drones, aircrafts and satellites. The era of satellite based EO began in 1959 with the launch of Explorer 7, and remains until today [32]. In fact, there are more than 2000 active EO satellites operated by Space Agencies, governmental institutions and commercial operators [11,33], resulting in an increased availability of information concerning the Earth condition and proprieties [34]. EO data are an example of a big data source that can be acquired at low cost, over long periods of time and used to comprehend the entire Earth system while addressing scientific challenges [35] such as climate change and global warming [36], ecological change and reduction impacts of habitat and biodiversity deterioration [37] and used to produce statistics and indicators that enable the quantification of SD [11,12]. The United Nations report [12] has demonstrated the viability of using EO data to produce official statistics, including SDGs statistics such as agricultural [38], urban and land planning [39] or food security indicators [40].
EO satellite imagery can be classified into two groups, based on the sensor used to capture images: the passive sensors receive emitted or reflected radiation by the Earth's surface, and the active sensors emit radiation and receive the echoes reflected or refracted by the Earth's surface [11]. Overall, EO sensors provide data at four different resolutions: spectral, spatial, radiometric and temporal. The spectral resolution is the ability to define/ distinguish wavelengths ranges of radiation; hence, different spectral bands provide a spectral signature for specific land cover types [11] such as soil [41], water [42] or buildings [43]. The spatial resolution refers to the area that each pixel represents on the surface, the radiometric resolution indicates the degree of light intensities the sensor is able to distinguish [44] and the temporal resolution is related to the revisit time, namely the frequency with which sensors cross a specific area on Earth. Besides the differences related to the type of EO sensors, the data provided by satellites can also be distinguished by the different orbits. The geostationary orbit means that satellites track the same area and the Low Earth orbit means that satellites track the surface as they orbit [11].
EO images can be used to identify characteristics of interest based on how images are presented and their inherent properties, such as in agriculture [45], forests [46], water [47] and urban areas [48]. Identifying such characteristics has been often seen as a classification problem which requires techniques to classify or group pixels, according to their spectral characteristics, as belonging to a class [48]. The study of Group on Earth Observations [10] has identified SDGs that are measurable, at some level, using EO data. Figure 2 presents SDGs that can already be measured and analysed based on EO data as SDG 2-No hunger, SDG 6-Clean Water and Sanitation, SDG 13-Climate Action, and SDG 14-Life Below Water.
Taking advantage of emerging developments within EO domain represents an accurate and reliable way to address the SDG indicators and targets and thus bridge the gap between developed and developing countries discrepancy on the quantity and quality of data [20]. The data from EO sources have been advocated by several international organisations and researchers, such as Holloway et al. [38] and Murthy et al. [14], as a mean of minimising costs compared to the conventional acquisition and monitoring of different environmental parameters over relevant scales, areas and time periods [11].
From Fig. 2, it can be depicted that EO can provide quite a large number of indicators for the SDG framework such as data on the condition of the atmosphere [49], oceans [50], crops [51], forests [52], climate [53], natural disasters [54], natural resources [55], urbanisation [56], biodiversity [57] and human conditions [58]. The two most important indicators are population distribution (I-1), and cities/infrastructure mapping (I-2) since they contribute to all the SDGs. On the other hand, the SDGs which benefit from all the EO indicators are the zero hunger (SDG 2), clean water and sanitation (SDG 6), climate action (SDG 13), life below water (SDG 14) and partnership for the goals (SDG 17). This view is supported by the Global Working Group on Big Data [59] and United Nations [12] that states that satellite imagery has significant potential to provide more timely information, minimising the number of surveys and offering more disaggregated data for informed decision making. As a consequence of the quantity of data generated by EO sources, the necessity to find methods to process and analyse this amount of data arises. The purpose is to transform the EO data into valuable information.

Earth observation using machine learning techniques
In the last decade, there have been some major contributions to a wide range of Earth Science applications, from analysing gases, soil, vegetation, climate and, more recently, to ocean [60,61]. Recent advances on Machine Learning (ML) field are creating unprecedent opportunities to evaluate and monitor policy decisions as well as humanitarian initiatives [62,63]. Despite the advantages of using ML techniques, it may require greater computational resources as well as an expert to interpret results. ML techniques can be classified into four groups: supervised, unsupervised, semi-supervised and reinforcement learning schemes. The major difference between supervised and unsupervised lies in the fact that the first one requires output values (classification) in the training dataset [64] where problems can be either as classification or regression techniques. In contrast, unsupervised learning techniques require only the input values in the training dataset since their purpose is to find hidden patterns in data and can be handled by clustering or dimension reduction techniques [65]. Semi-supervised learning combines aspects of supervised and unsupervised learning and requires a combination of data with and without classification [66]. Reinforcement learning aims to build systems that can learn from the interaction with the environment, using rewards and punishments rules [67,68].
The following sub-sections give an overview of the different techniques and methods pertaining to the use of ML in the scope of SD supported in EO data highlighting major findings and applications. This summary outlines the boundaries of research concerning the application of ML algorithms as well as their importance, relevance and potential to support further research towards the development of robust methodologies concerning universal applications. This overview takes into consideration the most recent research results as well as their relevance.

SDGs tackled with machine learning
ML is a subdomain of Artificial Intelligence, which according to Samuel [69] aims to provide to machines the ability to learn from data without being explicitly programmed. The study and development of algorithms plays a major role in ML, as it aims to build a model between inputs and outputs, based on the data and  [10] algorithms provided, to learn how to make decisions upon unseen information [70,71]. The popularity of ML is vast and increasingly applied to different subdomains, including Statistical Learning methods, Data Mining, Image Recognition, Natural Language Processing and Deep Learning [72].
A substantial number of ML algorithms have been used and described in the literature, performing a wide range of tasks in a variety of domains like Agriculture [73], Renewable Energies [74], Disasters [54], Climate [75], Construction [76], Human Living Conditions [58] and Health System [77]. Figure 3 presents the most relevant techniques applied to remote sensed data, grouped according to the four categories of supervised and unsupervised methods: classification, clustering, regression and dimension reduction.

Classification
A classification method belongs to supervised learning category, and it is applicable in cases where the overall aim is to accurately assign a datapoint to a class [78][79][80]. There is a broad range of classification methods as presented in Table 2, in the scope of SD, that clearly shows the impact and potential use of these techniques in conjunction with EO data.

Clustering
The clustering method belongs to unsupervised learning category, and it is appropriate when the purpose is to associate/divide datapoints into clusters [78,89]. Table 3 synthesises the findings within the scope of clustering methods used in combination with EO data to aid in the development of SDGs.

Regression
A regression method belongs to the same category as the classification method, supervised learning, and it is applicable when the aim is to predict/estimate a continuous output variable of a given datapoint [78,99]. There are several approaches, as presented in Table 4, in the scope of SD, that clearly show the impact and potential use of these techniques in conjunction with EO data.

Dimension reduction
Dimension reduction, similar to clustering method, belongs to the unsupervised learning category and typically follow two main approaches: Feature Selection (FS), applicable when there is the necessity to select fewer characteristics [111,112]; and Feature Extraction, when the information needs to be synthesised through transformation. The aim is to create a small set of features covering much of the details in the initial dataset [79,113,114]. Then, these features/characteristics can be fed into other algorithms or otherwise used as an end result [78]. Table 5 synthesises the finding within the scope of dimension reduction methods used in combination with EO data to aid in the development of SDGs.

Methodologies and techniques for EO imagery analysis
Pre-processing, post-processing and the seldom incorporation of qualitative information play a major role in the success of any data analysis approach and is found to vary significantly among researchers. As above mentioned, the  [78] majority of methods for processing EO data are based on ML algorithms, whether they are supervised or unsupervised [11]. However, besides the general problem category, the techniques can also be classified according to the approach used taking into consideration images analysis and their feature extraction: Sub-PB, PB, Super-PB and OB [129,130]. In Sub-PB, each pixel can have multiple classes [131,132]; in PB, it is only possible to have one class per pixel [133,134]; in Super-PB, the pixels are grouped based on homogeneity [135,136]; while in OB, the aim is to delineate readily usable objects from imagery or partitioning an image into objects [137,138]. Figure 4 illustrates the Sub-PB, PB, Super-PB and OB techniques.
In addition to those techniques, there are visual interpretation techniques conducted through direct operator analysis of characteristics from raw satellite images. Such techniques are used to extract visual characteristics including colour, form, size, pattern, texture and shadow from images [11]. The human abilities, however, should be explored/emulated to further enhance and automate ML algorithm-based image interpretation. Overall, several approaches are being used by different researchers that combine ML algorithms and pre-processing of data giving rise to different methodologies.

Empirical and semi-empirical modelling
Empirical and Semi-Empirical models are created based on data acquired from observations or experiences, which means that there are none or few assumptions on data analysis. There are many examples of the application of empirical and semi-empirical modelling, such the ones in Table 6:

Supervised classification techniques
The Supervised Classification requires a set of classified samples (sub-pixels, pixels or super-pixels) to train the models to understand each class' patterns. After training models should be able to categorise new samples or place those samples into classes [143]. Some applications of these approaches are presented on the following Table (7).

Unsupervised classification techniques
Unsupervised Classification techniques do not require any training data or prior knowledge, and their main goal The proposed method performed similar to SVM and RF in the classification of crops with similar phenology [57] Developed an efficient framework for multi-temporal crops classification [82] SDG 6 (Clean Water and Sanitation) Wetland The developed framework for coastal plain wetlands classification had high accuracy. [83]

SDG 8 (Decent Work and Economic Growth) Slavery
The approach was used to help to liberate slaves by mapping brick kilns. [58] SDG 11 (Sustainable Cities and Communities) Land use The approach based on CNN achieved an accuracy of ≅ 98% for land use and land cover analysis [84] The proposed approach confirmed its suitability for urban planning because it had a superior performance compared to the global one [56] Living conditions Deep learning demonstrated a high potential to map areas of deprived living conditions [85] Land cover The multivariate time series algorithm showed high accuracy for rare land cover classes [86] SDG 13 (Climate Action) Climate The model based on decision trees, and used to classify local climate zones, achieved a good performance [75] SDG 14 (Life Below Water) Marine habitat SVM and K-NN classifiers achieved an accuracy higher than 90% on mapping coastal marine habitat [50] SDG 15 (Life on Land) Land cover The approach used allowed to differentiate the hyperspectral subclasses from the classes [87] Forest Sentinel-2 is considered a powerful source of data for forest monitoring and mapping [52] RF was the best method to predict and map the area and volume of eucalyptus [88] is to group image pixels or sub-pixels into unlabelled classes [11]. Table 8 lists some recent examples regarding the application of unsupervised classification techniques.

Image segmentation object-based classification
The image segmentation OB classification is used to identify objects based on their proprieties or features. These techniques were developed to emulate the human visual interpretation. Some applications of OB techniques are presented in Table 9.
The success cases presented in Tables 2, 3, 4, 5 and in Sect. "Methodologies and techniques for EO imagery analysis", demonstrate that the contribution of ML is crucial towards the analysis of data provided by EO sources.
The synergy between EO and ML can be viewed as an important tool to support a wide variety of SDGs and fields at a global scale and enhance their level of implementation, effectiveness and efficiency. Some of the most common SDGs presented in this paper, which benefits from the synergy EO-ML are: SDG 11, 15 and 9; and the most common fields are Agriculture, Land Cover and Pollution.

Conclusions
Sustainability is an unavoidable aspect for the development of societies and countries; it leads to the development of SDGs and, hence, is crucial to the future of the planet. SDGs are unique as they cover issues that affect The results showed that FCM was superior to K-Means and Self-Organising Map for mineral favourability mapping [90] SDG 11 (Sustainable Cities and Communities) Land change The proposed approach based on K-Means, demonstrated better detection accuracies and visual performance for land cover and land change detection, compared to several methods [91] Seismic The method analysed was reliable and effective in the identification of sequences of earthquakes [92] Construction The proposed method used to segment individual buildings had a good performance with datasets acquired from densely built-up areas [76] Land cover The proposed clustering method outperformed the original approach for remote sensing segmentation in land cover classificatio [93] SDG 13 (Climate Action) Wildfires The presented algorithm for global burned area mapping was capable to adapt to different ecosystems and spatial resolution data [54] Geomorphology The proposed DBSCAN methodology for geomorphological analysis allowed the detection of movements of a rock glacier [94] Climate The techniques used such as K-Means and DBSCAN demonstrated their suitability for predicting climate types [53]

SDG 14 (Life Below Water) Sandbars
The proposed algorithm demonstrated a high potential to be used for the extraction of sandbars positions [95] SDG 15 (Life on Land) Soil degradation Assessment of spatial variability and mapping of soil properties provide an important link in identifying soil degradation spots [96] Agriculture Optimised kernel-based FCM gave more accurate agriculture crop maps when compared with the classical FCM and K-Means [97] SDG 17 (Partnerships for the Goals) Sustainability level The results obtained using Hierarchical Cluster Analysis showed that Sweden has the highest level of sustainability among the European countries; while, Greece, Bulgaria and Romania were the countries with the lowest performance [98] all communities and reaffirm the international commitment to eradicate poverty, hunger and inequalities to build a more sustainable, prosperous and safer planet for all humanity. This paper highlights the importance of monitoring the SD by means of EO and ML and enhances their fundamental role in pursuing those goals. Monitorisation aspects related to SD, such as poverty, nutrition, health conditions and inequalities have leveraged EO data collection methods. EO is possibly the most cost-effective technology as it is able to provide data at a global level and therefore enabling a global perspective of the SDGs. EO data plays a critical role in promoting equity among developed and developing countries since it grants worldwide data access despite their development level. EO data analysis, which often involves identifying features of interest within large amounts of information (Classification, Clustering, Regression or Dimension Reduction problems), gets even more powerful through the application of ML methods using different methodologies such as Empirical and Semi-Empirical modelling, Sub-PB, PB, Super-PB or even OB techniques. This extensive review looked at different ML categories to handle EO data to tackle different SDGs. It can be concluded that all ML categories can contribute to a wide variety of SDGs and fields-The Classification category covers the SDGs 2,6,8,11,13,14 and 15, and fields such as Agriculture, Land Use and Forests; the Clustering category covers the SDGs 2,7,9,11,13,14,15 and 17, and fields such as Construction, Natural Estimate the crop yield, at a pixel level, using ML proved to be an accurate approach [73] The results obtained from the comparison of methods showed that Boosted RT was the best to predict maize yield [ The best performance, to obtain the latent heat evaporation using a small dataset, was achieved by Kernel Ridge Regression, and using a large dataset, was achieved by Bagging RT [108] Grassland Vegetation indices acquired from Sentinel 2 have high potential concerning grasslands productivity, management, monitoring and conservation [109] Landslide Catchment map units and model selection are crucial for the performance of landslide susceptibility maps [110]  The proposed method demonstrates better results compared to other methods for land cover classification in almost all tests [122] Dimensionality Reduction was considered a key step in the land cover classification process [123] The experiments shown that the impervious surface extraction accuracy of Classification and Regression Tree was higher than Seperability and Thresholds algorithm [124] Land use FS with Classification Optimisation Score metric reduces the processing time and produces higher classification accuracy for land use and land cover classification using VHR data    Land use Tested Sub-PB and Super-PB methodologies to map green spaces. The results showed that Super-PB approach was better for dense urban, sub-urban and rural subsets. However, for lower-resolution images, the Sub-PB approach performed better for dense urban and sub-urban subsets [136] Land change Developed two CNN approaches: Early Fusion and Siamese Network to detect changes in pairs of images. Overall, the results proved that Siamese Network approach was the most accurate [144] SGD 12 (Responsible Consumption and Production) Consumption Proved that CNNs combined with high-resolution images represent a precise and cost-effective methodology to calculate consumption expenditure and wealth in developing countries [145] SDG 15 (Life on Land) Land cover Analysed 15 years of research on supervised classification methods and found that SVM was the most accurate among NN, RF and Decision Tree [146]  Land cover Used and compared three methods to classify ground vegetation covers using data acquired by IKONOS satellite. The comparison demonstrated that all methods are very accurate (more than 90% of accuracy); however, the two-step method achieved the best results [147] Land change Proposed an unsupervised method with an OB approach to improve the detection of changes using high-resolution images. This methodology achieved better results in comparison to other methods [148] SGD 15 (Life on Land) Invasive plants Developed an unsupervised method to detect and map invasive plants using RFs, which proved to be a successful approach [149] Landslide Compared an unsupervised PB and OB approach for landslide detection using VHR images and concluded that OB performed better than PB [150]  Land cover Compared four OB classifiers for the classification of a suburban area with data provided by Landsat-8 and proved that SVM had the best performance among all [151] Land use Proposed OB approach for urban land use classification using VHR images [152] SGD 15 (Life on Land) Land cover Tested the performance of PB and OB classification with a hyperspectral dataset and found that OB was better than PB approach [153] Land use Compared an OB and PB approach using aero photogrammetric images and the results showed that OB classifier performed better compared to PB [154] Disasters and Renewable Energy; the Regression category covers the SDGs 2, 3, 6, 7, 9, 11, 13, 14 and 15, and the fields Water Quality, Pollution and Freshwater; and the Dimension Reduction category covers the SDGs 3, 6, 7, 9, 11, 13 and 15, and the fields Land Cover, Electricity and Software. Thus, the overall findings confirm the significance of EO and ML in pursuing the goals of SD providing an overview of methods and techniques that sustain the achievement of SDGs. Lastly, the applicability and efficiency of specific ML methods used to analyse EO data, such as Random Forest (RF), Support Vector Machine (SVM) and Neural Network (NN), should be further explored to sustain a more consensual and reliable development/improvement of tools to support SDGs.