2.1. DL taxonomy for urban sprawl

The uses of DL algorithms for modelling urban sprawl indicators have evolved sufficiently in recent years to overcome the challenges of this phenomenon. Mastering the DL implementation process to model a given urban sprawl indicator requires knowledge of several elements. The different categories of these elements constitute the taxonomy of DL methods for urban sprawl applications. Figure 1 shows the diagram that summarizes this taxonomy and more specifically the domains involved.

Figure 1. Taxonomy of DL methods for urban sprawl applications.

The taxonomy of DL methods for urban sprawl applications in the selected study area (city) consists of four categories of elements including features, data source, DL methods (also known as DL models or algorithms), and targeted indicators. The urban sprawl features category groups together the various dependent variables that will be used to model the indicator targeted in the indicator category. Note that a target indicator in one problem may play the role of an input feature in another, and vice versa. Once the input features and target indicators have been defined for a given study, the data source category describes how the data has been collected to be more reliable. The fourth element of the taxonomy and the most important for this paper are types of DL algorithms that have been applied or are potentially applicable. In the following parts, we briefly go into more information about each issue.

2.2. Literature search method

To establish a knowledge base on a topic and locate articles to serve as a data source for additional examination, the literature search outlined in Figure 2 was implemented. The main stages in the literature search process are summarised in Figure 2 below according to preferred reporting items for systematic reviews and meta-analyses (PRISMA) methodology for systematic literature review (Mateo, Reference Mateo2020; Rethlefsen et al., Reference Rethlefsen, Kirtley, Waffenschmidt, Ayala, Moher, Page and Koffel2021).

Figure 2. Flow chart describing the different stages of our literature research process.

The first step of the research consisted of introducing the following keywords: “urban sprawl” OR “urban expansion” OR “urban growth” AND “deep learning.” Similar to other scientific database search queries, the review paper/book types, letters, or non-English documents were excluded using EXCLUDE instruction (Tékouabou et al., Reference Tékouabou, Chenal, Azmi, Toulni, Diop and Nikiforova2022a). So, the final research querapplications in urban sprawl feature predictiony in the Scopus database was structured as follows: TITLE-ABS-KEY ((“urban sprawl” OR “urban growth” OR “urban expansion”) AND “deep learning”) AND (EXCLUDE (DOCTYPE, “re”) OR EXCLUDE (DOCTYPE, “cr”) OR EXCLUDE (DOCTYPE, “er”) OR EXCLUDE (DOCTYPE, “sh”)) AND (EXCLUDE (LANGUAGE, “Chinese”) OR EXCLUDE (LANGUAGE, “Korean”)).

2.3. Literature outcome and filtering

After removing reviews, conference reviews, and letters, we obtained 281 raw articles and 183 after the initial filtering processes. The most pertinent publications—those discussing deep learning applications in urban sprawl feature prediction—were retained after key data about the chosen papers were obtained as .csv files and filtered (Tékouabou et al., Reference Tékouabou, Chenal, Azmi, Toulni, Diop and Nikiforova2022a, Reference Tekouabou, Diop, Azmi and Chenal2023a). Following a series of keyword-based filterings, data-analysis strategies, and topical mining with the ORANGE tool, we adjusted the search parameters to obtain relevant publications that either directly or indirectly reference the description and, consequently, the source of the related urban sprawl data (Tekouabou et al., Reference Tekouabou, Diop, Azmi, Jaligot and Chenal2022b). A few related articles that were not saved from the earlier stages but were thought to be pertinent were manually inserted. Finally, the 57 most relevant publications were crucial to the analyses in the following sections. The list of these papers is provided as Supplementary Material in an Excel file.

3.1. Terminology and salient characteristics

The uncontrolled and frequently unplanned growth or expansion of urban areas into neighbouring rural or undeveloped territory is referred to as urban sprawl (Sarkar and Chouhan, Reference Sarkar and Chouhan2019). The phenomenon of a town or urban area being defined by its outward expansion results in low-density development, wasteful use of land, and dispersion of both infrastructure and population (Baqa et al., Reference Baqa, Chen, Lu, Qureshi, Tariq, Wang, Jing, Hamza and Li2021). The patterns of land use, infrastructure, transportation, the environment and the general standard of living in an area can all be significantly impacted by urban sprawl. (Prayitno et al., Reference Prayitno, Sari, Hasyim and SW2020; Aithal et al., Reference Aithal, Vinay and Ramachandra2018). Hence figure 3, which best describes the key features’ categories that include: density, land use/cover, infrastructures, landscape/layout, environment and finally socio-economic and development.

Figure 3. Key features of urban sprawl categorised according (Tekouabou et al., Reference Tekouabou, Diop, Azmi, Jaligot and Chenal2022b).

3.2. Outlook from the literature review analysis

Usually, a literature review involves two major categories of analysis: bibliometric analysis and thematic trends (Tekouabou et al., Reference Tekouabou, Gherghina, Kameni, Filali and Gartoumi2023b). The bibliometric analysis is provided in the Appendix 1, and we will confine ourselves to analysing the thematic trends in this section (Tekouabou et al., Reference Tekouabou, Diop, Azmi and Chenal2023a). To highlight thematic trends, we used two types of visual analysis based on recent text mining techniques on Orange data mining (word cloud visualization in Figure 4) and Vosviewer (co-occurrence visualization in Figures 5 and 6).

Figure 4. Word cloud of keywords.

Figure 5. Network visualisation of the co-occurrence of the keywords plotted by VOSviewer.

Figure 6. Density visualisation of the co-occurrence of the keywords plotted by VOSviewer.

The illustration of Figure 4 shows the word cloud from a corpus consisting of the title, abstract and keywords of the most relevant papers. This figure was generated by applying text mining to the Orange data mining tool. On this figure, we can see that the words associated with the first rank of our searches are much better weighted (for example deep, learning, urban, growth, expansion, etc.).

Figures 4 and 6 show the visualization of the co-occurrence of keywords in the most relevant documents selected for the study by network and density, respectively. With a minimum threshold of two occurrences, there are seven main clusters of topical trends, each consisting of:

• • Cluster 1: Built-up area, classification, CNN, deeplabv3+, machine learning, urban land cover, and urban remote sensing.

• • Cluster 2: Aerial images, building extraction, semantic segmentation, transfer learning, urban growth, and urbanization.

• • Cluster 3: Change detection, convolutional neural network, deep learning, remote sensing, sentinel-1, and urban expansion.

• • Cluster 4: Change detection (cd), convolutional neural network, deep learning (dl), and object detection.

• • Cluster 5: Data fusion, sentinel-2, urban sprawl, and Yangze River economic.

• • Cluster 6: Cartography, historical maps, and road extraction.

• • Cluster 7: Cellular automata, land use change, and urban expansion simulation.

The overall insight from the co-occurrence could be summarized in three means points: First of all, The thematic focus of all these clusters strongly emphasises remote sensing, CNNs, and change detection techniques to analyze urban sprawl. There is also a focus on applying deep learning to diverse data types (aerial, satellite, and historical) to model and predict urban expansion and land use changes. Secondly, the methodological insights, the common thread across clusters is the use of CNNs and deep learning architectures for classification, segmentation, and change detection tasks. This highlights the growing importance of AI techniques in urban growth prediction, where different types of imagery (optical, radar) and historical data are fused. The third point is the applications with a focus on case studies like the Yangtze River Economic Zone reflect the real-world application of these methods in regions experiencing rapid urbanization, suggesting practical policy implications for managing urban sprawl.

Globally, all these clusters collectively point to a multi-disciplinary approach combining AI thanks to deep learning methods, remote sensing, and spatial modelling to predict and understand urban growth patterns. However, the absence of certain key themes from this analysis shows that there are still many challenges to be met, offering gaps in the application of deep learning to simulate urban sprawl indicators.

3.3. Urban sprawl challenges

The challenges posed by urban sprawl are as numerous as they are varied. As with its characteristics, there is not necessarily a classic categorization of these challenges in the literature. However, as shown in Figure 7, we could propose to group them into three main categories: socioeconomic challenges, environmental challenges, and infrastructural challenges.

Figure 7. Illustrating the main challenges of urban sprawl.

3.4. Urban sprawl issues in African cities context

Africa and Sub-Saharan Africa (SSA) in particular, has the highest urban growth rate, at around 3.6% per year (Yiran et al., Reference Yiran, Ablo, Asem and Owusu2020; Bocquier, Reference Bocquier2005). It is often considered the fastest urbanising region in the world with an urban population of around 472 million people at present and this figure will double over the next 25 years (Saghir and Santoro, Reference Saghir and Santoro2022). Forced or even surprised by this strong population growth, Africa’s main cities are heavily affected by the complex phenomenon of urban sprawl, which stems from rapid and above all unplanned urban growth (Forget et al., Reference Forget, Shimoni, Gilbert and Linard2021). In addition to demographic growth, urban sprawl in Africa is also characterised by indicators such as informal extensions to built-up areas often without adequate infrastructure development. The early development of areas with no regulations or basic services, the lack of formal housing are the consequent creation of shanty towns (Tekouabou et al., Reference Tekouabou, Diop, Azmi, Jaligot and Chenal2022b, Reference Tekouabou, Diop, Azmi and Chenal2023a). Concrete examples include Lagos in Nigeria, one of the fastest-growing cities in the world, where sprawl poses challenges in terms of waste management, transport infrastructure, and the provision of basic services (Yiran et al., Reference Yiran, Ablo, Asem and Owusu2020). Another example is the city of Nairobi in Kenya, which is expanding rapidly with informal developments on the outskirts, exacerbating inequalities, and putting pressure on natural resources and infrastructure (Yiran et al., Reference Yiran, Ablo, Asem and Owusu2020). Urban sprawl in Africa raises several major challenges including proper planning of the areas occupied once the population has settled in an anarchic manner, which leads to social breakdown and discontent, not to mention socioeconomic inequalities (Yiran et al., Reference Yiran, Ablo, Asem and Owusu2020). Another major challenge concerns infrastructure and basic public services which are non-existent and often difficult to implement once the population has settled (Tekouabou et al., Reference Tekouabou, Diop, Azmi and Chenal2023a). This brings with it de facto problems of transport, mobility, and security, as well as the emergence of small groups of bandits and the development of risky informal transport (Chenal and Le, Reference Chenal and Le2017). All these challenges are not without mentioning the environmental challenges arising from the destruction of natural ecosystems, deforestation around towns, and often the failure to create green spaces in occupied areas (Tekouabou et al., Reference Tekouabou, Diop, Azmi and Chenal2023a). Potential solutions to the problem of urban sprawl in Africa include developing urban planning services, facilitating procedures for registering land for legal sale, introducing affordable housing policies, and improving infrastructure. Finally, there is the funding of urban development research into the use of new techniques to strengthen the planning and monitoring of urban occupation in line with policies set in advance (Gartoumi and Tékouabou, Reference Gartoumi and Tékouabou2023).

3.5. The need for precise and forecast methods for modelling urban sprawl

The significant effects of urbanization on the environment, society, and economy need the development of forecast urban sprawl modelling methods (Shao et al., Reference Shao, Sumari, Portnov, Ujoh, Musakwa and Mandela2021). To properly understand, foresee, and manage these ramifications, accurate modeling is necessary. In urban planning, reliable insights from accurate modelling are essential for informed decision-making by urban planners, policymakers, and stakeholders (Purswani et al., Reference Purswani, Verma, Jayakumar, Khan and Pathak2022; Tope-Ajayi et al., Reference Tope-Ajayi, Adedeji, Adeofun and Awokola2016; Dinda et al., Reference Dinda, Das, Chatterjee and Ghosh2019; Arsanjani et al., Reference Arsanjani, Helbich, Kainz and Boloorani2013). These decisions pertain to land use, infrastructure development, and resource allocation, influencing the trajectory of urban growth (Hua and Gani, Reference Hua and Gani2023; Jat et al., Reference Jat, Garg and Khare2008). Resource management becomes critical in the face of urban sprawl which strains land, water, and energy. Predictive modelling identifies areas of potential resource overuse, guiding strategies for efficient utilization, and conservation (Sarkar and Chouhan, Reference Sarkar and Chouhan2019; Hua and Gani, Reference Hua and Gani2023; Osman et al., Reference Osman, Shaw and Kenawy2018). Infrastructure planning, encompassing transportation networks, and public services benefits from accurate modelling to meet future needs and optimize investments (Purswani et al., Reference Purswani, Verma, Jayakumar, Khan and Pathak2022). Environmental conservation relies on modelling to identify ecologically important areas and sensitive habitats, informing land use policies that balance development with conservation efforts (Sarkar and Chouhan, Reference Sarkar and Chouhan2019; Tope-Ajayi et al., Reference Tope-Ajayi, Adedeji, Adeofun and Awokola2016). Understanding the environmental impact of sprawl aids in formulating targeted strategies for sustainable growth, climate change mitigation, and improving air and water quality (Abudu et al., Reference Abudu, Echima and Andogah2019; Osman et al., Reference Osman, Shaw and Kenawy2018; Dinda et al., Reference Dinda, Das, Chatterjee and Ghosh2019). Economic planning involves forecasting economic consequences and enhancing resilience, while policy evaluation ensures the effectiveness of land use policies (Sarkar and Chouhan, Reference Sarkar and Chouhan2019; Osman et al., Reference Osman, Shaw and Kenawy2018; Dinda et al., Reference Dinda, Das, Chatterjee and Ghosh2019). Ultimately, the goal of urban sprawl modelling is to foster sustainable development, creating livable, resilient, and environmentally conscious urban environments through informed decision-making and thoughtful planning (Sarkar and Chouhan, Reference Sarkar and Chouhan2019; Osman et al., Reference Osman, Shaw and Kenawy2018; Dinda et al., Reference Dinda, Das, Chatterjee and Ghosh2019; Jayasinghe et al., Reference Jayasinghe, Raghavan and Yonezawa2021).

4.1. Conventional methods for predicting urban sprawl

A variety of methods have been employed to simulate and forecast the growth of urban areas through the use of traditional urban sprawl modelling techniques (Padilla et al., Reference Padilla, Netto and Da Silva2020). Among the prominent types of conventional urban sprawl modelling methods are cellular automata (CA) and statistical models. (Belinga and El Haziti, Reference Belinga and El Haziti2023; Adegun et al., Reference Adegun, Viriri and Tapamo2023).

The application of statistical approaches in urban sprawl features modelling is generally from a quantitative perspective (Padmanaban et al., Reference Padmanaban, Bhowmik, Cabral, Zamyatin, Almegdadi and Wang2017) for leveraging historical data. Various statistical techniques have been used to understand the factors influencing urban expansion (Tope-Ajayi et al., Reference Tope-Ajayi, Adedeji, Adeofun and Awokola2016). These models often focus on relationships among socioeconomic variables (Khan and Sudheer, Reference Khan and Sudheer2022), land use patterns (Brown et al., Reference Brown, Pijanowski and Duh2000), and demographic trends (Congedo and Macchi, Reference Congedo and Macchi2015). Notable statistical-based techniques within this category include: regression models, logistic regression, Markov chain-based models, and spatial econometric-based models (Jiang et al., Reference Jiang, Chen, Zha, Zhang and Cui2022; Qiang et al., Reference Qiang, Lin, Zhu, Wu and Lee2021).

Spatially explicit models known as (CA) split the study region into distinct cells or pixels; every single one which might be in any number of states (e.g., urban, agricultural, and unoccupied) (Belinga and El Haziti, Reference Belinga and El Haziti2023; Sarkar and Chouhan, Reference Sarkar and Chouhan2019). These models mimic land-use changes according to preset criteria that dictate how cells move from one condition to another (Belinga and El Haziti, Reference Belinga and El Haziti2023; Huang et al., Reference Huang, Huang, Liao, Liang and Tian2018). CA offers a dynamic representation of urban sprawl and can capture interactions between neighbouring cells. Key features of CA-based models include transition rules (Arsanjani et al., Reference Arsanjani, Helbich, Kainz and Boloorani2013; Cilliers et al., Reference Cilliers, van Vuuren and van Heerden2021), Neighborhood influence (Akın and Erdoğan, Reference Akın and Erdoğan2020), time steps (Cilliers et al., Reference Cilliers, van Vuuren and van Heerden2021), and calibration and validation (Akın and Erdoğan, Reference Akın and Erdoğan2020; Cilliers et al., Reference Cilliers, van Vuuren and van Heerden2021).

Table 1 summarises the published papers according to the methods they use. CA and statistical models have both made significant contributions to our understanding of the dynamics and causes of urban sprawl. They do have certain drawbacks, though, such as oversimplifying intricate procedures or making assumptions that might not hold in all circumstances (Belinga and El Haziti, Reference Belinga and El Haziti2023). Contemporary urban sprawl modelling frequently combines these more established methods with more recent ones, such as ML (Purswani et al., Reference Purswani, Verma, Jayakumar, Khan and Pathak2022) and remote sensing (Tope-Ajayi et al., Reference Tope-Ajayi, Adedeji, Adeofun and Awokola2016; Padmanaban et al., Reference Padmanaban, Bhowmik, Cabral, Zamyatin, Almegdadi and Wang2017), to improve precision and offer a more thorough comprehension of patterns of urban expansion.

Table 1. Traditional methods used by article

4.2. DL techniques for simulating urban sprawl

To improve the accuracy and predictive power of urban sprawl models, various DL-based techniques have often been used in recent years (Tekouabou et al., Reference Tekouabou, Diop, Azmi and Chenal2023a). These methods, which make use of deep neural networks (DNN) can learn from enormous datasets, capture complex spatial correlations, and produce more accurate forecasts regarding urban expansion (Belinga and El Haziti, Reference Belinga and El Haziti2023). Tekouabou et al. (Reference Tekouabou, Diop, Azmi, Jaligot and Chenal2022b) presented potential ML techniques suited for urban sprawl’s features prediction according to urban form.

4.2.4. Transfer learning and generative models

For certain urban sprawl modeling tasks, the pre-trained DL models such as CNNs trained on extensive picture datasets like Imagenet (LeCun et al., Reference LeCun, Bengio and Hinton2015) can be adjusted. This method makes use of the pre-trained model’s expertise and customizes it for the urban environment (Belinga and El Haziti, Reference Belinga and El Haziti2023). However, using past data, generative models such as generative adversarial networks (GANs) and variational autoencoders (VAEs) can produce artificial urban growth scenarios (Man and Chahl, Reference Man and Chahl2022). These scenarios help analyze the effects of various policy initiatives and investigate possible future patterns of urban expansion. Table 2 groups the browsed articles according to the DL methods they use.

Table 2. Deep learning methods use by article read