Introduction

Changes in the surface of the globe have significant consequences for the global radiation balance and energy flows, contributing to changes in biogeochemical cycles and changes in hydrological cycles, and affecting ecological balances and complexity. It is generally held that understanding land use and land cover change is essential to the public discourse on sustainability. It has been widely reported that the globe has been subject to a significant number of variations over the last decades [1,2,3]. These have frequently been correlated with the rise in deforestation, desertification, and land degradation observed in large parts of the area, and there has been a need for constant monitoring and oversight [4, 5].

However, despite its greatness to scientists and policy makers confronting the challenges of environmental change, land use and land cover change is still not well underpinned. Remote sensing offers cost-effective multispectral and multi-temporal information that is useful for the analysis and monitoring of land use models and trials and for the creation of land use and land cover data sets in a geographic information system context that provides a versatile environment for storing, analysing, and viewing digital information that is important for the identification of change and database evolution [6, 7].

Urban growth is actually a big category of land use and land use change, as it is related to a rise in populations and their financial systems [8]. More fundamentally, it has significant environmental consequences, such as the degradation of water and air quality [9]. It also has socioeconomic outcomes, including economic gaps, community disintegration, and improved infrastructure charges [10].

Remote sensing information has confirmed that it is important to observe and identify urban transformation, and to provide the necessary information for future progress. In the past, researchers and planners relied on visual aerial photo analysis to develop land exploitation plans [11]. Nevertheless, land use maps can now be made more appropriate, inexpensive, and more accurate with the growth of remote sensing devices and the availability of the latest high-resolution digital technology [12]. Urbanised growth, driven by increasing population, economic development, and technological progress, has become an international phenomenon. Such development is often seen as the energy of a territorial economy [10, 13]. However, it has had unconstructive effects on the environment such as biodiversity deprivation, soil erosion, hydrological disturbances, water and solid contamination, and global warming. Monitoring and modelling urban spatial development are important for environmental sustainability and urban planning [6, 14].

The latest environmental and biological changes in the context of widespread and enhanced social inequality, along with the prevalence of poor living conditions, have had an effect on the pattern of disease spread. These environmental, economic, and political developments increase epidemiological uncertainty, favouring the introduction of new pathogens and the re-emergence of old infectious diseases, historically linked to rural communities and potentially existing in new settings [8, 9]. These concerns have prompted health-related organisations and, in addition, international institutions such as the World Health Organization to take action. A review of the scientific knowledge gained from the application of this approach to the study of endemic diseases, such as the coronavirus disease, is continuously carrying out. Analysis aims to evaluate the potential for the use of this technology as a tool for the study and management of the most prevalent and most dangerous diseases.

Recognition of changes is the method of identifying differences in land cover by monitoring them at various times [15]. Different methods of recognition of changes are usually compared to the most significant results of detection of changes for a specific function [15, 16]. The most commonly used methods of change detection are image differentiation, post-classification, transformation vector testing, and principal component analysis [17]. Studies confirm that these methods have been able to effectively recognise changes in diverse land cover through multi-temporal satellite images for different purposes [18, 19]. Furthermore, the research looked at methods for the detection of functional change for urban change [14, 20].

Satellite Imagery

Satellite imagery refers to the assignment of globe imagery from detectors and sensors placed on sophisticated satellites in orbit around the globe [6, 21]. Satellite images provide significant data that can be used in a number of remote sensing applications, such as meteorology, cartography, urban change recognition, and agricultural inspection. Campbell and Wynne [22] have identified and categorised the satellite images by four different types of image resolution:

  1. 1.

    Spatial resolution: The spatial resolution of a satellite picture refers to the dimension in square meter of the globe, calculated on the ground, which is covered by a picture pixel. Nevertheless, the most general metric used to illustrate spatial resolution is the metre, which explains the length of the ground on each side of the square image pixel [23]. The spatial resolution is influenced by the altitude of the satellite orbit around the globe, and also the type of sensors positioned on satellite. The higher the spatial resolution of the image is, the lower the surface area a pixel represents. As a result, a satellite image of elevated spatial resolution allows for a particular land vision given that even small objects are captured and then visible [24].

  2. 2.

    Spectral resolution: The spectral resolution of a satellite image is identified as the number of spectral bands it contains and the size of wavelength variation calculated in nanometres that each satellite image band captures [25]. This means that the spectral information extracted from each satellite may change. Most satellites can provide images with at least four channels where, apart from the standard red-green-blue (RGB) channels, there is also a near-infrared channel involved. For example, the IKONOS satellite provides 4-channel satellite images, while the LANDSAT-7 satellite gets 7-channel images. Moreover, even the wavelength interval of the same spectral channel may vary from sensing element to sensing element, and this means that the same spectral channel may contain slightly different information when captured from different sensing elements. Table 1 shows the wavelength and resolution for each specific band (http://web.pdx.edu/~nauna/resources/10_BandCombinations.htm). For example, the green channel of the QuickBird satellite consists of 520−600 nm, while the green channel of the WorldView-2 satellite consists of 510−580 nm. In addition, Table 2 shows the combination bands and corresponding frequencies (http://maps.unomaha.edu/Peterson/gis/notes/RS2.htm).

  3. 3.

    Temporal resolution: The temporal resolution of a satellite image refers to the time, normally counted in days, required by a satellite to perform a complete circle around the globe and return to the same position as before with the objective to accumulate picture of the identical surface zone [26]. This case of resolution depends mainly on the orbit of the satellite around the earth.

  4. 4.

    Radiometric resolution: The radiometric resolution of the satellite image refers to the degrees of brightness of the sensing elements that can be captured by the satellite. The radiometric resolution is described in bits and the most typical measurements are 8, 11, 12, and 16 bits. The range of values that the pixels of a satellite image can achieve depends on the radiometric resolution of the sensor used to capture the image [27]. In particular, given that the resolution of the radiometric image is n bits, the intensity of the image pixels is within the range (0, 2n–1). As a result, the pixel size of the 8-bit resolution image is within the range of 0–255, while the pixel values of the 16-bit resolution image are within the range of 0–65535 [28]. Figure 1 provides a global overview of the resolution types of the different optical satellite systems [29].

Table 1 Wavelength (μm) and resolution (m) bands of Landsat 7
Full size table
Table 2 Landsat TM band combinations
Full size table
Fig. 1
figure 1

An overview of spectral, spatial, temporal, and radiometric resolution of different optical satellite systems [26]

Full size image

Land Use Change Detection and Land Covering

Land use and land cover change play a key role in global change and have important implications for many international policy developments. Over the last few decades, population growth and urbanisation have led to a dramatic transition from grassland and forest to cropland and residential areas. As a result, land use and land cover change have been a major concern for its results in terms of the water cycle, biological diversity, soils and their fertility, and water and air quality [30]. Careful consideration is shown by local and regional land managers and policy makers to take decisions that balance positive growth and bad with the idea of maintaining environmental resources and enhancing socioeconomic well-being. In order to understand how land use and land cover change make and interact with global world schemes, it is necessary to provide information on what changes occur, when and where they seem, the values at which they occur, and the natural and social drivers of these changes [31].

Satellite remote sensing data may provide useful information on the geographical dispersion of land use and land cover change, along with price and time savings for information gathering over traditional inventories and studies across large regions [32]. In addition, data on remote sensing may provide information on the temporal trends and spatial distributions of urbanised regions involved in understanding, modelling, and designing land change. Figure 2 shows the main devices used for remote sensing imaging [33]. Remote sensing change detection is the process of identifying differences in the status of an object or phenomenon by observing it at different times. Remotely sensed data has become a major source of information for various change detection application programmes in recent decades, due to repetitive data acquisition and digital formats that are suitable for computer processing [34]. Although most of the change detecting research studies have focused on land use and land cover change, recent work has used change detection techniques in a number of applications, such as forestry [35] and vegetation modification, forest fire [36], wetland change, and urban change [37].

Fig. 2
figure 2

The main devices used for remote sensing imagery [33]

Full size image

Urban Change Detection

Land cover change and land use are key elements of global environmental change and have significant impacts on many aspects of ecosystem processes and functions at local, regional, and global scales. Over the last few decades, agricultural expansion and intensification, deforestation, and urbanisation have progressively affected more than half of the Earth’s unfrozen surface [38]. A number of published studies have suggested that land cover change and land use have significant impacts on biodiversity, water production and quality, climate, soil status, food protection, and human well-being [39,40,41]. Although urbanised regions comprise only a small part of the world’s land area, urbanisation is widely viewed as the most intensive and important land cover change and land use type [42]. Land transformations from forest and other natural land to urbanised land are typically permanent and may lead to a number of ecosystem-related problems, such as impaired air and water quality, the spread of invasive species, and habitat fragmentation [43]. With the rapid migration of population to urban areas, urbanisation processes have accelerated over the last decades. Monitoring past and recent urban scale and urbanisation operations is of vital importance for the assessment of ecosystem services and for the promotion of policy making [44].

Urbanisation is one of the most important forms of global change in both developing and developed countries. Urbanisation is caused by the growth of population and economy, as well as by the concentration of human population in large metropolitan areas. Inappropriate land preparation and land management cover changes associated may be associated to. The urban development may have major impacts on climate and ecosystems, including degradation of water and air quality, loss of farmland and forests, and the socioeconomic effects of economic disparities, social fragmentation, and monetary value of infrastructure [45]. In addition, urban areas have higher solar radiation absorption and thermal capacity and conductivity than other land-based types. As a result, urban areas tend to have higher temperatures than adjacent agricultural areas due to impervious surfaces such as roads and buildings. This thermal difference leads to the development of urban thermal islands [46]. Monitoring and representing the urban development needs robust techniques and methods. Traditional study and mapping methods do not have the capacity to provide the necessary data in a timely and cost-effective manner. In recent decades, studies have shown the potential for remote sensing modification detection techniques to provide accurate and timely data on urbanised land change [47, 48]. Mas [49] found that post-classification is the most successful method of urban change detection, with information on various soil moisture and vegetation phenology circumstances. In addition, Li and Yeh [50] concluded that multi-date principal component analysis (PCA) and categorization could be successfully applied to determine rapid urbanised expansion in the Chinese river. Principal component analysis is the foundation of a multivariate analysis of data based on projection techniques. The most significant application of principal component analysis is the implementation of a multivariate record board as a small set of variables with the aim of viewing trends, bounds, groups, and outliers [51]. Principal component analysis is a very flexible device and allows the study of datasets that may include, for example, multi-collinearity, misplaced values, and categorical records, in addition to incorrect data [52]. The objective is to detect and express significant data records as a set of primary indices or main components [51, 53]. Anees et al. [54] used remote sensing with geographic information systems (GIS) change detection techniques to map the spatial dynamics of urban land modifications.

The main challenge for the categorization of urbanised land cover is to accurately identify the relative contribution of reflection from a number of materials in an urbanised environment. Furthermore, consideration needs to be given for changes that are not linked to urbanised land cover change, such as seasonal vegetation variations. In addition, some works, including other spatial data, have been investigated in order to overcome the sub-pixel mixing problem and improve per-pixel compartmentalisation by using textural [55] and contextual [56] information.

Change Detection

Detection of changes in computer vision can generally be identified as a problem with the identification of the character and magnitude of the changes between the two compared images that describe the corresponding scene and were obtained at different time periods. Depending on the objectives of each task, some changes are fundamental and require further analysis and modelling, while others are superficial or improper and should be discarded. Detection of change can be a challenging task as it affects various elements, such as temporal, spectral and spatial, restraint, radiometric resolution, and atmospheric conditions [57]. In an effort to overcome the difficulties and provide accurate change in the detection of outcomes, different changes to detecting skills have been prepared. Authors grouped change detection techniques based on different points of view and grouped them into pre-classification or post-classification methods [58] and pixel-based or object-based methods [59]. Pixel-based change detection methods consider the image pixel to be the basic unit of image analysis and change. These techniques, which can also be considered as pre-classification techniques due to the fact that they are employed on raw data, are most often used in the literature on change detection methods. Several pixel-based change detection methods have been developed and their advantages and disadvantages have been analysed in an attempt to improve the accuracy of the change detection [60, 61].

Famous pixel-based change detection techniques include image differentiation, image rationing, picture regression, background subtraction, principal component analysis, and change vector analysis (CVA). Change vector analysis is used to recognise spectral changes involving two identical scenes that have been produced at different times. Change vector analysis is limited to two bands for each image. The change vector in the spectral dual-dimensional space is estimated for each pixel. It explains the transformation of one attribute across a variety of events (variables) between two limits of time as a vector within the space of variables [62]. Fundamentally, a vector with a direction element and a magnitude can be expressed. The magnitude constituent expresses the quantity of change as the direction element indicates the type of change. Picture differentiation calculates the absolute differences between the pixel values of the grayscale image and is the easiest way to execute, but no information on the type of change is provided. An overview of the use of image differentiation for land cover change detection was provided in [63]. Image rationing calculates the ratio of grayscale pixel measurements and it has the reward that it is better able to handle calibration errors [64] but no complete change data matrix is provided. Picture regression attempts to adjust the pixel values of the objective image to those of the reference image by regression analysis. It accounts for variations in mean and variance between pixel measurements for different dates and thus reduces the adverse effects caused by atmospheric conditions and sun angles [65]. Unfortunately, it is difficult to determine precise regression functions due to local luminance fluctuations present in images. Background subtraction [66] attempts to find the background information from the photos compared and then deducts it from the photos in order to obtain change information. Although simple to implement, the background subtraction is subject to low accuracy [67].

Principal component analysis applies dimensionality reduction to the withdrawal of information redundancy between multispectral bands and the punctuation of different information. Principal component analysis was used in [68] as a distance metric to determine the magnitude of change in the value of the pixel when the Bayes principle was used to distinguish pixels in two classes, “change” and “no change”. The disadvantages of the principal component analysis change detection methods are the difficulty in interpreting and labelling the results of the principal component analysis and the lack of extracted data on the nature of the changes [69].

In contrast to previous pixel-based change detection techniques, the change vector analysis can provide information on the type of change. Change vector analysis uses all the available multispectral image data, interprets each pixel as a vector, and differentiates on vectors. As a result, the Euclidean vector distance identifies the magnitude of the change, while the vector angle provides information on the type of change. The disadvantage of changing vector analysis change detection method is that it requires accurate radiometric correction for atmospheric effects to be mitigated and such precise radiometric correction is not possible [19]. All pixel-based change detection methods suffer from two major disadvantages or limitations that have led to research into object-based change detection techniques. The first drawback is that a threshold is necessary to distinguish between pixels that belonging to the “change” class from those that belonging to the “no change” class. Selecting the optimum threshold is a problematic procedure and pixel-based techniques are usually subject to mis-detection or over-detection [70]. This is because the threshold is too weak to allow too many surfaces of change, while the threshold is too high to expel surfaces of change. The other limitation of pixel-based techniques is the fact that they do not take into account the spatial links between adjacent pixels [71]. This exclusion from the spatial context leads to the removal of primary information, which would greatly contribute to accurate results in the detection of changes. The special subcategory of pixel-based change detection techniques is based on texture analysis methods. These methods are used at the window/block level, implying that the statistical characteristics are calculated over the imbrications windows that move across the compared images. As a result, these techniques take into account the spatial setting of pixels and they may be more noise-proof than other pixel-based techniques.

Change detection solutions for these algorithms are based on the rating of similarity metrics between the windows being compared. A common similarity metric used in several texture analysis supported change detection methodologies is a correlation that can be used either in the time domain in the form of normalised cross-correlation (NCC) [72] or in the frequency domain by transforming images using Fourier and then analysing the index spectrum (phase correlation (PC)) [73]. NCC is a main mathematical instrument for signal and image processing for character matching, correspondence analysis, signal tracking, and object identification [74, 75]. In order to improve its real-time and efficient performance, it was suggested that digital standardised cross-correlation be implemented by some fast algorithms and hardware structures due to its high computational complexity [76, 77].

In addition, the Bhattacharyya distance is used to compare pixels inside the imbrication windows [78]. In digital image processing, an estimate of the theoretical distance involves two normal distributions of spectral categories, which operates as a higher boundary on the possibility of miscalculation in a Bayesian approximation of the acceptable classification.

Texture-based methods are subject to the same limitation as different pixel-based techniques, which is the demand for the best threshold that distinguishes pixels that correspond to true changes from all other pixels. In an attempt to overcome the limitations of pixel-based change detection techniques, object-based methods have been developed. Object-based methods can usually be identified as post-classification methods due to the fact that the images are commonly fragmented into regions or categories before any change detection method is used. The unit for analysis in object-based change detection techniques is an image object with valuable data, such as texture, shape, and spatial relations with adjacent objects, providing for the use of spatial settings [79].

Picture partition is at the centre of the object-based change detection methods and different supervised or unsupervised partition techniques are used with varying results. Most object-based change detection techniques can be classified in either field-based (similarity of pixels) or boundary-based (discontinuity of pixels) or edge-based methods [80]. The expectation-maximization (EM) algorithms [81], the artificial neural networks (ANNs) [82], and the support vector machine (SVM) classifiers [7, 83] are image partition algorithms that are usually used in object-based change detecting methodological analyses. The expectation-maximization algorithm is a methodology for achieving the maximum probability assessment of hidden variables. It does this by first approximating the values for hidden variables, then optimising the prototype, and then reiterating the two steps of the union. It is an efficient and common methodology and is most commonly used for density evaluation through missing records, such as clustering algorithms such as the Gaussian mixture model.

In addition, support vector machines are a kind of machine learning classifier, possibly one of the most common types of classifiers. Support vector machines are particularly useful for numerical prediction, categorization, and sample detection tasks [12, 21]. Support vector machines run through the representation of decision limits between records points, aiming at a decision boundary that most excellently divides the information points into categories. The objective when using a support vector machine is that the decision boundary between the points is as large as possible so that the distance between the specified data point and the boundary contour is maximized.

Change detection methods based on the EM algorithm achieve higher accuracy than other methods but they require a priori knowledge of the joint probability class [81]. The artificial neural networks approach can provide significant change detection results, particularly when categories are not normally distributed, but it needs a lot of preparation time and it is sensitive to the amount of training data [84]. Artificial neural networks are seen as nonlinear statistical data modelling tools where complex relationships involving inputs and outputs are designed or patterns are found. Data flowing through the network affects the structure of artificial neural networks because the neural network varies on the basis of input and output. Artificial neural networks are deep learning frameworks capable of forming identification and machine learning. They are part of the broader domain of artificial intelligence (AI) technology [85].

Finally, the advantage of support vector machine classifiers is that they do not make assumptions about the underlying data distribution and process the task of detecting change as a dual classification problem. The support vector machine classifier learns from the training records and the optimum threshold is automatically defined [86] from the spectral characteristics with the aim of classifying the pixels in the “no change” or “change” categories. Other machine learning algorithms used to monitor categorization, classification, and change detection include fuzzy logic [87], random forest [88], or genetic programming [89]. Fuzzy logic is a valued logic structure that may have accurate variable values in any actual number between 0 and 1. The fuzzy logic algorithm makes it easier to resolve a problem after taking into account all existing records. After that, the most likely result for the specified input is taken. Common phenomena, such as environmental change, rarely have crisp limits and therefore cannot be approached in the conventional method by means of dual logic (weak or not: 0 or 1) [90]. As a result of fuzzy logic, units are allowed to have a partial association with a given category, which is allocated through bias formulas. Among the fuzzy relational calculus, a numerical foundation is built to treat all types of cases where the uncertainty of the status of the environment cannot be ignored.

Lately, random forest algorithm has become particularly famous in the field of remote sensing, and is rather quick compared to other machine learning methodologies [91]. The concise description of the random forest algorithm is derived from the denomination. Rather than employing the forecast of an individual decision tree, the algorithm will result in a large number of decision trees being assembled. The “random” fraction of the name comes from the expression “bootstrap aggregating”, or “bagging”. This means that every tree within the forest can only learn to train on some division of the complete training dataset. The components of the training records for each tree left hidden are believed to be “out-of-bag” for precision evaluation. Uncertainty also helps to having the choice which variables of the input attribute are observed at each node in each decision tree. Once all the trees are robust for random division of the training records, by means of an arbitrary set of attribute variables for each node, the assembly of all the trees is used to provide the final forecast [92].

Genetic algorithms (GA) are the ideal option for heuristic unsupervised classification. Heuristic unsupervised classification is based on the creation of some mathematical designs and, subsequently, the improvement of a predefined indicator to determine the cluster information and the centroids mechanically [93]. Heuristic optimization procedures are considered to be a reproducible, accurate, and time-efficient method of repeatedly categorising remote sensing images. Previously, genetic algorithms have been successfully approved for image processing [94] and image identification for specific uses such as therapeutic treatment [95].

Object-based techniques that rely on image segmentation suffer from over-segmentation or under-segmentation miscalculations. Such miscalculations produce objects that do not accurately represent real-world objects [96]. Under-segmentation creates larger objects that hide more than one object in the real world, and they may contain mixed categories of objects, whereas over-segmentation creates finer components of real-life objects that need to be merged to make current objects more pragmatic [97]. Unfortunately, the miscalculations of the used image classification process are disseminated in the final change map, which reduces the accuracy of the most recent change detection results [98]. In order to improve the results of the change detection, the classification of individual images must be as accurate as possible.

Conclusion

Apparent urban growth, driven by rising population, economic expansion, and technological progress, has become an international trend. Such growth is often seen as the strength of a local economy. However, it has had unconstructive effects on the environment, such as biodiversity loss, soil erosion, hydrological disturbances, water and solid contamination, and global warming. Screening and modelling of urban spatial expansion are essential for ecological sustainability and urban planning. Due to the ability to deal with nonlinear and multifaceted phenomena, the integration of analytical techniques and methods with remote sensing has enhanced the performance of image classification for the disjointed and varied landscape in the urban environment. The other component is to differentiate spatial urban expansion in the practical zone, and cellular phases by using three methodologies: urban land change planning, site metrics survey, and moving windows examination.