Optimization of road route alignment: a systematic literature review with meta analysis

Abstract

This systematic literature review (SLR) integrates Geographic Information Systems (GIS), deep learning, and Multi-Criteria Decision Making (MCDM) to enhance road route optimization, crucial for global infrastructure development. This SLR aims to identify existing research trends, methodologies, research gaps and propose a generalized framework for streamlining the road route optimization process. The review addresses three key research questions: RQ-1. The application of deep learning for Land Use and Land Cover (LULC) classification, RQ-2. The use of MCDM techniques in road route alignment and RQ-3. Techniques for optimizing road route alignment. Utilizing PRISMA, we assessed 370 papers, selected 132 through full-text evaluation, and added 25 via. snowball sampling, totalling 157 records for analysis. The results reveal trends in current research, geographical distribution and the evolution of methodologies. It is found that Deep learning techniques significantly improve LULC classification accuracy, while MCDM techniques enable a holistic approach to road route alignment by incorporating diverse factors. The proposed generalized framework outlines a systematic approach encompassing problem definition, criteria selection, data preparation, deep learning-based LULC classification, MCDM and Least Cost Path analysis for road route alignment. This work uniquely identifies research trends, methodologies, and gaps in road route optimization, addressing three specific research questions (RQ-1 to RQ-3) on deep learning (LULC classification), MCDM techniques, and route alignment optimization. This work also highlights the scope for integrating emerging technologies, enhancing MCDM approaches, promoting cross-disciplinary collaboration, addressing data availability and quality, conducting case studies, emphasizing sustainability, resilience and focusing on global and regional contexts. This SLR will surely contribute to the development of efficient, sustainable and equitable road route optimization strategies for better infrastructure planning and worldwide development.

1 Introduction

Geographic Information System (GIS) has become an essential tool in modern geoinformatics and geo-planning. Their integration with advanced technologies like AI and deep learning is revolutionizing how we approach complex challenges such as road alignment and linear infrastructure development. Traditionally, road construction relied heavily on manual surveys and measurements. This approach, while effective in its time, often neglected critical factors such as environmental, economic, engineering and social considerations.

In today’s world, where road networks are continuously expanding to support growth and meet sustainable development goals, there’s an increasing need for AI-driven road route optimization. This approach leverages the power of GIS combined with the analytical capabilities of deep learning to enhance the efficiency and effectiveness of road development projects.

Roads are the backbone of transportation infrastructure, playing a vital role in connecting communities, facilitating trade, and driving economic growth (World Bank 2018). Recognizing the importance of roads, governments worldwide invest heavily in road construction and maintenance (OECD 2019). In the United States, the Biden administration has put forward a proposal for a $2 trillion infrastructure initiative aimed at considerable enhancements to the nation’s infrastructure, including major investments in roads, bridges, and highways (The White House 2021). Similarly, India had ambitious plans to expand its road network, with the government having aimed to construct 65,000 km of national highways by 2022. This project had been expected to incur a cost of Rs. 5.35 lakh crore (approximately US$ 741.51 billion) https://www.fhwa.dot.gov/policy/otps/pubs/impacts/. These investments not only improve connectivity and reduce travel times but also create employment opportunities (Federal Highway Administration 2021). However, the development of road infrastructure must also consider sustainability and environmental concerns, such as greenhouse gas emissions and habitat loss (United Nations Environment Programme 2020).

Failure to integrate GIS with deep learning in this domain can lead to persistent delays, inefficiency, violations of pre-set constraints and suboptimal routing decisions. Given the high demand for sustainable road development, addressing these challenges is more urgent than ever.

Moreover, the methodologies and frameworks developed for optimizing road routes using GIS and deep learning can be applied to other linear engineering problems such as railway route alignment, selection for oil and gas pipelines, metro route selection, and evacuation route selection. This review aims to develop a comprehensive framework for road route optimization, leveraging the combined strengths of GIS, deep learning and optimization to address the complex demands of modern infrastructure development. This approach is not only about improving road construction but is also a step towards more intelligent, efficient and sustainable infrastructure planning and development.

The 2023 study by F. Jiang et al. presents a comprehensive framework for sustainable urban road alignment planning, focusing on the identification and evaluation of key factors (Jiang et al. 2023). This study categorizes factors into demands such as traffic and local development and constraints including environmental, engineering, cost and social aspects. The framework employs digital tools and spatial analysis methods like GIS, digital image processing and digital twin technology to cohesively format these factors for effective planning. Experts use MCDM methods to assess the importance of each factor. Furthermore, the study introduces a least-cost wide path analysis for road alignment which considers road width and accommodates both new road construction and the expansion of existing roads.

In a related domain, the 2023 review paper by Fei Li et al. proposes integrating remote sensing with machine learning techniques to enhance urban sustainability (Li et al. 2023). This integration aims to make urban sustainability more explicable and efficient. While the literature review by F. Jiang et al. contributes significantly to the field, it also opens avenues for future research, particularly in integrating AI technologies like deep learning and further exploring MCDM techniques (Jiang et al. 2023). Similarly, the work of Fei Li et al. highlights the potential of combining machine learning with remote sensing to address urban sustainability challenges (Li et al. 2023).

The existing literature on route alignment primarily discusses various factors involved in developing new routes and the methodologies for route selection. However, with the rapid advancements in artificial intelligence, especially in deep learning, there emerges a need to explore the role of deep learning in route alignment. Additionally, while MCDM techniques play a crucial role in the road route alignment process, there is a noticeable lack of comprehensive literature analyzing MCDM techniques specifically for road route alignment.

Recent advancements in Multi-Criteria Decision-Making (MCDM) techniques have significantly improved decision support capabilities in complex, uncertain, and data-scarce environments. Notably, the Grey MARCOS method (Measurement of Alternatives and Ranking according to Compromise Solution) and various fuzzy-based MCDM techniques have gained traction due to their ability to incorporate vagueness and ambiguity in real-world decision-making. For example, Akbulut (2025b) used the Grey MARCOS model to evaluate financial performance in the Turkish insurance sector, while Demir (2025) applied fuzzy MCDM for security risk assessment and countermeasure selection in smart cities. Similarly, Akbulut (2025a) provided a comparative analysis using Grey PSI and Grey MARCOS, further reinforcing the robustness of such hybrid models under imprecise conditions.

The methodological sophistication of MCDM techniques has grown with the integration of fuzzy sets, Z-numbers, intuitionistic fuzzy environments, and spherical fuzzy sets. For instance, Hussain and Ali (2025) introduced an advanced intuitionistic fuzzy Z-number-based decision-making model to evaluate ideological and political education for sustainable development. Meanwhile, Badi et al. (2025) proposed a fuzzy MCDM framework for optimizing sustainable logistics in Free Trade Zones. Recent methodological contributions such as the Parsimonious Spherical Fuzzy Analytic Hierarchy Process (Moslem 2025) and the Z-number extension of the Parsimonious Best Worst Method (Moslem 2025) have further enhanced the granularity and expressiveness of decision models. Additional works by Ullah et al. (2025) and Hussain et al. (2025) demonstrated applications of intuitionistic fuzzy aggregation and fuzzy group decision-making in domains such as urban transport safety and digital tool selection.

While these models demonstrate considerable potential in modeling subjective and imprecise judgments in diverse domains, their application in spatially-intensive infrastructure problems like road route alignment remains limited. In contrast, the Analytic Network Process (ANP) provides a clear advantage in modeling the intricate interdependencies and feedback loops among diverse spatial, socio-economic, and environmental factors that characterize road infrastructure planning. For example, slope, land use, hydrology, population density, and construction cost often exhibit nonlinear interdependencies, which are effectively captured by ANP. Moreover, unlike models that treat criteria as isolated, ANP facilitates a network-based structure, enabling robust prioritization under interconnected influences. Hence, while this review acknowledges the methodological value of recent fuzzy and hybrid MCDM models, it adopts ANP as the core decision-support tool for its contextual suitability in road alignment tasks. The integrated framework proposed herein combines ANP with GIS-based spatial analysis and deep learning-based land classification, offering a holistic approach for modern road route optimization.

Firstly, the research investigates how deep learning, a type of advanced technology in artificial intelligence, is used to study LULC Classification. Deep learning is good at understanding complex data and it can be helpful in identifying different types of land and how they are used.

Secondly, it looks at MCDM techniques. MCDM helps in making good decisions about where to place roads by considering many different factors. This study explores how MCDM is used in practice. Finally, the SLR examines techniques for optimizing road route alignment. Through the exploration of these three areas, the research seeks to offer comprehensive and current insight into the planning of road routes and the examination of land use in an intelligent and well-informed manner. This SLR includes insights from general observations of paper distributions, application fields, deep learning and MCDM approaches, along with road route optimization methods.

The remainder of the paper is structured to ensure clarity and ease of navigation, facilitating a thorough understanding of the systematic review conducted. Section 2 presents a detailed description of the methodology adopted, outlining the systematic literature search procedures, databases utilized, and criteria applied for the inclusion and exclusion of studies. In Section 3, the results derived from this systematic review are presented comprehensively, including general observations, a proposed generalized framework for route optimization, sustainability criteria, advancements in LULC classification through deep learning techniques (Esmaeili et al. 2023), application of various MCDM methods, and a critical analysis of algorithms for optimal route generation. Section 4 offers a focused discussion interpreting the significance of these findings, highlighting current trends, existing research gaps, and potential implications for practical applications and future studies. Finally, Section 5 summarizes key conclusions, emphasizing contributions of this review to both academia and industry, while suggesting directions for subsequent research.

2 Methodology

This study sets out to develop a comprehensive framework for the integration of deep learning in LULC analysis for optimizing road route alignment, utilizing MCDM techniques. This framework aims to enhance road route alignment projects by promoting sustainability. To achieve this, a SLR was performed using the PRISMA method, a well-acknowledged approach for structuring systematic reviews and meta-analyse. This systematic review enables an in-depth and broad understanding of the current research that merges deep learning with LULC and examines MCDM and road route optimization techniques. It identifies the existing state of knowledge, the methods employed and the potential research gaps that could be filled in future studies. The application of the PRISMA method guarantees that the review process is both reproducible and transparent, thereby reinforcing the study’s reliability and facilitating the verification and future update of its conclusions with new findings. The visual representation of the PRISMA Methodology is depicted in Fig. 1. The review process was structured into three stages: planning, conducting the review and reporting the findings.

Fig. 1

PRISMA diagram

To ensure that the review would offer valuable insights into the present opportunities and challenges, the planning phase included the definition of the research question, objectives, as well as the criteria for what would be included and excluded. The academic search engine provided by COEP Technological University Pune which offers access to various databases including Google Scholar, Science Direct and IEEE Xplore was utilized to gather related publications. The selection of IEEE Xplore, Google Scholar and Science Direct as primary search platforms was strategic due to their extensive access to a multitude of databases, facilitating comprehensive searches across diverse resources such as conference papers, journal articles, books, and more. A search query was executed on November 8, 2023. This SLR did not restrict the publication date in the search criteria to encompass as broad a range of relevant literature as possible. For this SLR, 3 research questions were formulated and specific search queries corresponding to these research questions were developed with the details provided in Table 1, queries were deployed for keyword-based searches, ensuring precision in data retrieval.

Table 1 Rephrased research questions for the SLR

The field of research focusing on optimizing road routes using deep learning and MCDM techniques is still developing. This makes it difficult to locate studies that perfectly align with this specific topic. The initial results of our search as outlined in Table 1, are summarized in Table 2. These initial findings are considered preliminary; they consist of raw data collected before applying any criteria for including or excluding studies.

Table 2 The results of the research queries in different databases

During our search in these databases, we refined the results by examining the titles and keywords to select those most relevant to our study. Additionally, we filtered the search results based on the type of content, such as books, book chapters, conference papers and research articles. We only selected research articles, conference papers and journal papers resulting in a total of 370 records

During the review phase, the initial step consisted of examining the abstracts of the collected literature and applying predetermined inclusion and exclusion criteria. From the initial corpus of 370 records, this process resulted in the identification of 133 primary records. It was noted that among these, 17 records were duplicates.

Subsequently, an exhaustive examination of the full texts of these 133 latures was undertaken, adhering to the inclusion and exclusion criteria to further refine the selection. This process resulted in the identification of 132 latures suitable for the primary study.

To augment the scope of relevant literature, the backward snowballing technique was employed on the selected 132 studies. This method involves reviewing the references within the primary studies to identify additional relevant works. Through this technique, 25 more studies were discovered, culminating in a total of 157 primary studies for comprehensive analysis.

The 157 selected studies were carefully analyzed and reviewed, focusing on five key aspects: a) Approach b) Dataset c) Limitations d) Advantages e) Accuracy f) Criteria Considered. In the final phase of reporting, a systematic review of these 157 studies was presented. This comprehensive analysis aimed to provide a clear overview of the current state of research in the field, identifying trends, gaps and presenting a framework for optimizing road route alignment through the application of deep learning.

Table 3 Inclusion and exclusion criteria

3 Results

3.1 General observations

All 157 research articles were categorized by country-wise distribution and it was found that the majority of the articles originated from India (24.06%), followed by China (15.04%), Turkey (4.51%) and the USA (3.01%). The country-wise publication distribution is visually represented in the word cloud shown in Fig. 2.

Fig. 2

Word cloud of countrywise publication

The research articles are classified into 3 groups according to the research questions to get a detailed understanding of each domain. Figure 3 shows the year-wise publications of the articles related to the road route alignment, it can be seen that the study covers articles from 1999 till 2023, also it can be seen that there is a drop in publication in 2022. Figure 4 shows the Year wise publication count for MCDM Techniques and it can be seen that there is an increase in publications from the year 2013 till 2021, from 2021 there is a decline in new publications and in 2023 there are no articles related to MCDM Techniques in the domain of route optimization specifically. Figure 5 shows the year-wise distribution of LULC with deep learning articles ranging from 2015 till 2023, It can be observed that there is an increase in the number of publications in 2023 as compared to previous years. This shows the trend in the field and growth is expected to continue in 2024 (Fig. 6).

Fig. 3

Route route alignment year wise publication

Fig. 4

MCDM techniques year wise publication

Fig. 5

LULC using deep learning year wise publication

Fig. 6

Generalized framework flowchart

3.2 Generalized framework

This study aims to provide a base to guide further studies regarding road route planning and optimization through the application of deep learning and MCDM techniques. This section aims to give an overview of the process, further sections describe individual components of the process in detail. This involves a threefold approach: identifying relevant factors and translating them into clear terms for road alignment planning (Vilke et al. 2018; Żabicki and Gardziejczyk 2020a, b), assessing these factors (Liu et al. 2022) and creating algorithms to derive optimal alignments (Vázquez-Méndez et al. 2021). Inspired by the Malczewski (1999) and Yakar and Celik (2014), the general framework of road route alignment can be outlined as follows:

• 1. Defining the endpoints and the problem at hand, specifically focusing on the source and destination between which the road needs to be built.

• 2. Identifying the criteria to consider such as environmental, social, economic and engineering factors which are crucial in road construction. These criteria may vary based on the study’s nature and should be referenced from existing literature.

• 3. Preparing and collecting the dataset relevant to the defined criteria, which may involve downloading geospatial datasets for the area under consideration. The dataset should include classes related to the defined criteria to create an overlay map.

• 4. Preprocessing the data to ensure consistency in format and resolution and to eliminate issues such as clouds and shadows for better classification. This involves converting vector data to raster format for ease of processing, considering that raster data structures allow for a more analytical approach in continuous space study, despite their disadvantages in accuracy and data storage (Eastman 2003).

• 5. Performing LULC classification using various deep learning techniques for image segmentation. The classes used for segmentation depend on the initially defined criteria.

• 6. Identifying and assigning ranks to the criteria, which could be derived from literature or determined by field experts. These ranks are crucial for the application of the MCDM techniques.

• 7. Applying MCDM techniques to identify the weights of the criteria based on the ranks provided by the field experts. These weights are crucial for the subsequent steps of the framework.

• 8. Apply the assigned weights to the raster data, creating a weighted raster that represents all the criteria. This step combines the LULC classification results with the criteria weights, preparing the data for the least cost path analysis (LCPA).

• 9. Perform least cost path analysis (LCPA) on the weighted raster to find the optimal route between source and destination.

• 10. Generate multiple candidate routes by varying the criteria weights in the LCPA step. This will provide a set of potential routes that optimize different aspects of the problem.

• 11. Apply MCDM techniques to evaluate routes based on the predefined criteria. This step provides a comprehensive evaluation of the generated routes from multiple perspectives.

• 12. Conduct sensitivity analysis to assess the robustness of the MCDM results by varying the criteria weights and observing how the rankings of the candidate routes change. If the rankings remain relatively stable, it indicates that the results are robust and less sensitive to changes in the criteria weights.

• 13. Select the final optimal route based on the MCDM rankings and sensitivity analysis results. The route that consistently performs well and is less sensitive to changes in criteria weights can be considered the most suitable choice.

Throughout this structured approach, the importance of maintaining data consistency in terms of format and resolution is emphasized to ease the processing steps, as highlighted in Fig. 7 depicting the raster and vector data formats (Żabicki and Gardziejczyk 2020a). Also, overlaying different map layers to normalize values into a single raster is essential, as illustrated in the process and Fig. 8, ensuring that multiple routes can be generated and evaluated by assigning different weights to criteria, enabling the selection of the best possible route according to specific project requirements and criteria. Figure 6 gives flowchart for generalized framework.

Fig. 7

Vector and raster storage (Żabicki and Gardziejczyk 2020a)

Fig. 8

Overlay map (Żabicki and Gardziejczyk 2020a)

3.3 Sustainability criteria

Sustainable practices in road construction are crucial for balancing environmental protection, economic efficiency, and societal well-being (Stern et al. 2021). Prioritizing construction on stable, less ecologically sensitive lands minimizes environmental impacts such as erosion and habitat disruption, aligning with sustainability goals to preserve natural resources and support community health (Dawson 2005).

Study (Jiang et al. 2023) provides a comprehensive list of criteria that can be considered during road development including traffic factors, economical factors, engineering factors, environmental factors and social factors; interested readers are encouraged to consult it for more information. Table 4 summarizes different factors identified across various studies in the literature.

Table 4 Comprehensive list of criteria found in literature

3.4 Land use land cover classification (LULC)

Land Use (LU) and Land Cover (LC) are critical components for understanding the spatial variations and the impacts of human activities on Earth’s surface. Land Cover refers to the natural state of the Earth’s surface, encompassing elements like soil, water, and vegetation. In contrast, Land Use denotes the alterations made to the land through human actions, such as deforestation, urbanization, and the development of built-up areas, along with natural phenomena like droughts and floods (Stern et al. 2021; Wang et al. 2023). Multiple researchers have tried the extraction of a single class to focus on specific problems like (Akhtarmanesh et al. 2023). To obtain accurate results, many researchers have performed the fusion of images of different resolutions (Mahdipour et al. 2024).

The study and classification of LULC changes are vital for remote sensing (Marzvan et al. 2021; Felegari et al. 2023), providing valuable insights through the extraction, processing, and classification of spectral signatures. Over recent decades, significant efforts have been directed towards automating LULC classification, a task that has become increasingly feasible due to advancements in remote sensing technologies (Safari et al. 2024). These advancements enable the analysis of large datasets, facilitating image classification, processing, and predictions of future changes (Abdollahi et al. 2020; Digra et al. 2022).

As per the study Fig. 9, reveals a total number of DL and ML techniques found in paper year-wise, this shows a growing trend of DL methods in the domain of LULC. Initially, traditional methods such as Support Vector Machines (SVM), Maximum Likelihood Estimation and Decision Trees dominated the field. These techniques offered timeliness and good repeatability, especially when compared to manual, visual interpretation methods. However, their classification accuracy tends to diminish when data or study areas change (Xie et al. 2022; Zhao et al. 2023).

Fig. 9

LULC techniques year-wise count

As the volume of data and the complexity of its analysis have grown, the machine learning (ML) community has explored various algorithms for image classification within LULC. Yet, the burgeoning data and the advent of new technologies and datasets have introduced classification challenges that transcend the capabilities of conventional machine-learning approaches. In this context, the availability of a wide array of socioeconomic data provides essential information for urban development planning and analysis (Abdollahi et al. 2020; Digra et al. 2022). The research shows that LULC data is mainly used in urban areas, as indicated by Fig. 10, which shows the domain-wise count of the paper.

Fig. 10

LULC domain-wise paper count

Deep Learning (DL) has emerged as a powerful tool capable of addressing these challenges. Unlike traditional algorithms, deep learning algorithms excel in processing large-scale data, identifying intricate patterns, and uncovering regularities (Bidwe et al. 2024), Bidwe et al. (2025). This ability significantly enhances the accuracy and efficiency of LULC classification, making deep learning a pivotal technology in the field (Abdollahi et al. 2020; Xie et al. 2022).

Table 5 demonstrates the evolutionary trajectory of both deep learning and traditional machine learning approaches in LULC classification from 2015 to 2024. The analysis reveals a pronounced shift toward deep learning methodologies, with Convolutional Neural Networks (CNNs) emerging as the predominant architecture, accounting for 19 implementations in the reviewed literature. Deep Neural Networks (DNNs) constitute the second most prevalent approach with 8 documented implementations. This prominence of CNNs can be attributed to their robust capability in spatial feature extraction from remote sensing data. The literature evidences significant diversification in architectural approaches, including region-based CNNs (R-CNN), transfer learning-enhanced CNNs, and various CNN variants. The emergence of Generative Adversarial Networks (GANs) and specialized Artificial Neural Networks (ANNs) in recent publications indicates the field’s progression toward more sophisticated architectures (Bidwe et al. 2022). While traditional machine learning algorithms maintain their relevance, particularly Support Vector Machines (SVM) with 12 implementations and Random Forests with 10 implementations, the trend clearly favors deep learning approaches. This transition from conventional machine learning to deep learning architectures reflects the growing complexity of LULC classification challenges and the superior capability of deep learning models in handling large-scale, multidimensional remote sensing data.

Table 5 Deep learning and machine learning methods year wise count (2015–2024)

LULC classification begins with downloading the dataset of the application area, Fig. 11 shows the most widely used data sources while performing LULC classification. The quality of data is important for better accuracy (Aryal et al. 2023). In pre-processing the input data is prepared for reducing its dimensionality, clipping raster images, image re-sampling, buffering and geo-referencing, denoise, synchronising, eliminating irrelevant information from the data, fusion of data (Digra et al. 2022), study (Song and Woodcock 2003) states that atmosphere is the primary source of noise for the accurate measurement also it talks about the algorithms used to remove atmospheric noise from the data. Figure 12 shows the generalized framework of LULC.

Fig. 11

LULC data sources paper count

Fig. 12

LULC generalize framework (Digra et al. 2022)

Once the data is ready one can proceed towards classification, The research in urban land use and land cover classification has advanced significantly, with various studies employing deep learning and other innovative methodologies to improve accuracy and efficiency. Lv et al. (2015) demonstrated that Deep Belief Networks (DBN) excel in contextual mapping using RADARSAT-2 data, achieving 81.74% overall accuracy and outperforming traditional methods like SVM and NN in the Great Toronto Area. Alhassan et al. (2020) developed a deep learning framework using Landsat 5/7 imagery for Manitoba, Canada, achieving up to 90.46% accuracy with the help of generative adversarial networks. Carranza-Garcí et al. (2019) highlighted the effectiveness of Convolutional Neural Networks (CNN) in classifying LULC with superior accuracy over SVM and random forests. Henry et al. (2019) introduced an automated mapping technique using deep neural networks, focusing on high-resolution aerial imagery for LULC mapping. Ienco et al. (2019) presented the TWINNS architecture, enhancing land cover classification by integrating Sentinel-1 and Sentinel-2 data, showing improved performance across study sites. Bhosle and Musande (2019) reported high accuracies in LULC classification and crop identification using a CNN model on hyperspectral images. Similar studies are being carried out by Farmonov et al. (2024). MirhoseiniNejad et al. (2024) have used Deep Learning techniques for crop yield prediction. Garg et al. (2019) introduced the mUnet architecture, offering improved performance over Unet and FCN for LULC classification in Karachi. Mu et al. (2019) achieved a 93.1% accuracy with a self-adaptive cellular-based deep learning approach for LULC change prediction. Zhang et al. (2020) proposed the SS-JDL method for autonomous scale selection in CNN input, enhancing remote sensing image classification. Rousset et al. (2021) assessed deep learning versus XGBoost for LULC classification in New Caledonia, showing deep learning’s superiority in land use detection. Naushad et al. (2021) explored deep transfer learning, significantly improving classification accuracy with the EuroSAT dataset. Zhu et al. (2021) introduced a hybrid UNet-ConvLSTM method for land use classification, achieving up to 84.4% accuracy. Luo and Ji (2022) proposed DACST for enhanced land cover classification across different times and locations. Alshari et al. (2023) combined ANN with RF for improved LULC classification accuracy using multispectral images. Boonpook et al. (2023) developed LoopNet, an advanced deep learning algorithm for land use classification using Landsat 8 imagery. Kumar and Gorai (2023) optimized a DCNN with SGDM for automatic classification in the Jharia coalfield, demonstrating remarkable performance. Pande and Banerjee (2023) unveiled a hybrid model blending 3D and 2D CNN layers for multimodal remote sensing image analysis. Usmani et al. (2023) integrated OpenStreetMap’s information with remote sensing data for global-scale segmentation, achieving high accuracies in urban feature analysis. These studies collectively showcase the evolution of LULC classification methodologies, emphasizing the potential of integrating diverse data sources and deep learning techniques for environmental monitoring and urban planning, The following are the most widely used deep learning techniques for LULC classification:

3.4.1 Stacked autoencoder (SAEs)

An Autoencoder (AE) is recognized as the fundamental component of a Stacked Autoencoder (SAE). An AE consists of a visible layer with $i$ inputs, a hidden layer with $N$ units, and a reconstruction layer with $i$ units. The training process encompasses two phases:

• 1. Encoding: Mapping input $\textbf{x} \in \mathbb {R}^i$ to hidden representation $\textbf{h} \in \mathbb {R}^N$.

• 2. Decoding: Mapping hidden representation $\textbf{h}$ back to reconstruction $\textbf{y} \in \mathbb {R}^i$.

The encoding and decoding processes are mathematically expressed as Eqs. 1 and 2.

$$\begin{aligned} & \textbf{h} = f\left( \textbf{w}_h \textbf{x} + b_h\right) ,\end{aligned}$$

(1)

$$\begin{aligned} & \textbf{y} = f\left( \textbf{w}_y \textbf{x} + b_y\right) , \end{aligned}$$

(2)

where $\textbf{w}_h$ denote the weights from input to hidden layer and $\textbf{w}_y$ from hidden to output layer respectively, $b_h$ and $b_y$ are the biases, and $f(\cdot )$ is an activation function. The estimation of the reconstruction error is achieved by minimizing the squared Euclidean distance, $\Vert \textbf{x} - \textbf{y}\Vert _2^2$.

SAEs are built by stacking multiple AEs in a way that the output of one layer serves as the input of the next. This architecture allows for the use of SAEs as spectral classifiers (Fig. 13).

Fig. 13

Stacked autoencoder (Zang et al. 2021)

3.4.2 Deep belief network (DBNs)

DBN (Hinton et al. 2006) has shown strong capability for unsupervised feature learning within the computer vision domain. A Restricted Boltzmann Machine (RBM) is a layer-wise training model essential to the Deep Belief Network (DBN) architecture. An RBM comprises a network with visible units $\textbf{v} = \{0,1\}^d$ and hidden units $\textbf{h} = \{0,1\}^L$. The energy of a joint configuration ($\textbf{v}, \textbf{h}$) is given by Eq. 3:

$$\begin{aligned} & E(\textbf{v}, \textbf{h}; \theta ) = -\sum _{i=1}^d b_i v_i - \sum _{j=1}^L a_j h_j - \sum _{i=1}^d \sum _{j=1}^L w_{ij} v_i h_j\nonumber \\ & = -\textbf{b}^\textbf{T} \textbf{v} - \textbf{a}^\textbf{T} \textbf{h} - \textbf{v}^\textbf{T} \textbf{w h}, \end{aligned}$$

(3)

where $\theta =\{b_i, a_j, w_{ij}\}$ includes the biases $b_i$ and $a_j$ for visible and hidden units respectively, and $w_{ij}$ represents the weight between them.

DBNs are formed by stacking RBMs to enhance feature representation. The final layer can be a logistic regression layer, making DBNs suitable for classification tasks (Fig. 14).

Fig. 14

Deep belief network

3.4.3 Convolutional neural network (CNNs)

CNNs are inspired by the visual system and utilize local connections to capture 2-D spatial features. They consist of convolution layers, pooling layers and fully connected layers (Fig. 15). The key components are:

• 1. Convolutional Layers: Input is convolved with filters to produce feature maps. For input $\textbf{X}$ and filter $j$, the output is 4:

  $$\begin{aligned} \textbf{y}_j = \sum _{i=1}^d f\left( \textbf{x}_i * \textbf{w}_j + b_j\right) , \end{aligned}$$

  (4)

  where $f(\cdot )$ is an activation function, e.g., ReLU defined as $\sigma (\textbf{x}) = \max (0, \textbf{x})$.

• 2. Pooling Layers: Reduce the spatial size of feature maps to make the representation more abstract, using operations like average pooling.

• 3. Fully Connected Layers Flatten the feature maps and transform them into a feature vector for classification or regression tasks.

Fig. 15

Convolutional neural network (Zhao et al. 2023)

3.4.4 Recurrent neural network (RNNs)

Recurrent Neural Networks (RNNs) (Chung et al. 2014) process sequences by maintaining a hidden state that captures temporal information. The hidden state $\textbf{h}^{\langle t \rangle }$ at time $t$ is updated as 5:

$$\begin{aligned} \textbf{h}^{<t>} = {\left\{ \begin{array}{ll} 0 & \text {if } t=0, \\ f_1\left( \textbf{h}^{<t-1>}, \textbf{x}^{<t>}\right) & \text {otherwise}, \end{array}\right. } \end{aligned}$$

(5)

with $\textbf{w}$ and $\textbf{u}$ being the weight matrices for the current input and previous state respectively, and $f_1$ a nonlinear function (Fig. 16).

Fig. 16

Recurrent neural network (Zhao et al. 2023)

3.4.5 Generative adversarial network (GANs)

Generative Adversarial Networks (GANs) (Voulodimos and Doulamis 2020) consist of a generator $G$ creating fake inputs and a discriminator $D$ distinguishing between real and fake inputs. The training aims to solve Eq. 6:

$$\begin{aligned} \min _G \max _D V(D, G) = E_{\textbf{x} \sim p(\textbf{x})}[\log (D(\textbf{x}))] + E_{\textbf{z} \sim p(\textbf{z})}[\log (1-D(G(\textbf{z})))] \end{aligned}$$

(6)

GANs have been adapted for tasks like classification by incorporating auxiliary classifiers to enable multi-class label prediction (Fig. 17).

Fig. 17

Generative adversarial network (Zhao et al. 2023)

3.5 Multi-criteria decision making (MCDM

The complexity and challenges of selecting a suitable route alignment for roadways and highways are significant, as this task involves balancing a wide array of criteria—some of which may complement each other, while others may compete. This decision-making process is crucial in determining the final alignment of a highway, relying on a comprehensive assessment of various factors (Salah Sadek 1999). The utilization of MCDM techniques has become a cornerstone in this context, offering a way to handle multiple criteria simultaneously, thereby guiding decision-makers towards the most advantageous decision alternatives (Ahmed and Asmael 2015). Figure 18 shows the application domain in which MCDM techniques are used according to the literature and it shows that the transportation sector is the one where MCDM techniques are mostly used.

Fig. 18

MCDM application found in literature

Multi-criteria analysis has gained widespread adoption in transport project studies, especially those commissioned by public bodies, and is a prevalent tool in assessing EU-funded research projects (Delle Site and Filippi 2009). This methodology’s application ranges from selecting railway routes, as demonstrated by Ali A. Darvishsefat et al. in their proposal of a railway route between Rasht and Anzali (Darvishsefan et al. 2004), to evaluating road alignments through the integration of the Analytic Hierarchy Process (AHP) and GIS technologies by researchers like Tae-Ho, Tae-Ho et al. (2008).

Various studies illustrate the multifaceted approach to infrastructure planning using GIS and MCDM techniques. For instance, Mario De Luca et al. conducted a multi-criteria analysis in a GIS environment to identify preferential corridors for a high-speed line in Italy (De Luca et al. 2012), while S.Loganathan et al. optimized highway alignment using GIS and soft computing tools, underscoring the potential for enhancing route selection and environmental conservation (Loganathan and Elangovan 2013). Similarly, Hala A. Effat et al. employed MCDM techniques to design and assess highway route alternatives in the Sinai Peninsula, Egypt, integrating environmental and economic factors (Effat and Hassan 2013).

The breadth of applications for these methodologies is vast, from optimizing natural gas transmission pipelines, as explored by Yildirim et al. (2013), to urban planning efforts like selecting the best metro route in Baghdad using GIS, AHP, and TOPSIS by Ahmed and Asmael (2015). The effectiveness of GIS and MCDM approaches extends to selecting pipeline routes with minimal environmental impact (Abudu and Williams 2015), identifying suitable sites for highway construction (Rajadurai and Johnny 2015), and finding economic and optimal paths for connecting towns (Suleiman et al. 2015).

Integrating GIS with AHP has proven to be an efficient and cost-effective method for road alignment selection, as shown in studies by Sunusi et al. (2015) and others, which advocate for the wider adoption of these technologies to improve accuracy and sustainability in route planning (Gitau and Mundia 2017). The application of these methods in addressing complex infrastructure challenges, such as navigating around natural hazards or optimizing route alignments for environmental and economic benefits, is well-documented across various geographical contexts and project types (Nayana Padmani and Sudath 2017; Singh and Singh 2017; Cruz-Chávez et al. 2020; Sameer et al. 2021; Jiang et al. 2022).

These examples collectively highlight the dynamic and effective nature of combining GIS and MCDM techniques in infrastructure planning and development. By leveraging such integrated approaches, decision-makers can significantly enhance the efficiency, connectivity, and safety of transportation networks while minimizing environmental impacts and meeting the diverse needs of urban and rural communities. The word cloud presented in Fig. 19 illustrates the various MCDM techniques employed in the literature. Comparative analysis of some commonly used MCDM techniques is given in Table 6 and those are described as follows:

Fig. 19

Word cloud of MCDM techniques

Table 6 Comprehensive analysis of MCDM techniques

3.5.1 Analytical hierarchy process (AHP)

The Analytic Hierarchy Process (AHP) method, introduced by Saaty (1984), involves calculating the importance of each criterion using a scale ranging from 1 to 9, where 1 signifies equal importance, 3 moderate importance, 5 strong importance, 7 very strong importance, and 9 represents extreme importance.

Step 1: The pairwise comparison matrix includes the scales $(a_{ij})$ which is used to determine the importance of each criterion.

$$\begin{aligned} \begin{array}{c|cccc} A & {\textrm{Criteria}}\, 1 & {\textrm{Criteria}}\, 2 & \ldots & {\textrm{Criteria}}\, n \\ \hline {\textrm{Criteria}}\, 1 & a_{11} & a_{12} & \ldots & a_{1 n} \\ {\textrm{Criteria}}\, 2 & a_{21} & a_{22} & \ldots & a_{2 n} \\ \ldots & \ldots & \ldots & \ldots & \ldots \\ {\textrm{Criteria}}\, n & a_{n 1} & a_{n 2} & \ldots & a_{n n} \end{array} \end{aligned}$$

Step 2: A normalized matrix is generated by dividing each element by the sum of its column. The formula is given in Eq. 7

$$\begin{aligned} a_{ij}^\prime = \frac{a_{ij}}{\sum _{i=1}^n a_{ij}} \end{aligned}$$

(7)

Step 3: The average of the sum refers to the weights of each criterion. The formula is given in Eq. 8

$$\begin{aligned} w_i = \left( \frac{1}{n} \right) \sum _{j=1}^n a_{ij}^\prime , \quad (i, j = 1,2,3, \dots , n) \end{aligned}$$

(8)

Step 4: Saaty (1994) proposed a Consistency Index (CI) calculation to decide whether the comparisons are consistent. The formula is given in Eq. 9

$$\begin{aligned} CI = \frac{\lambda _{\max } - n}{n-1} \end{aligned}$$

(9)

Step 5: CI calculation requires determining $\lambda _{\max }$ (the principal eigenvalue) with the Eq. 10:

$$\begin{aligned} \lambda _{\max } = \frac{1}{n} \sum _{i=1}^n \left[ \frac{\sum _{j=1}^n a_{ij} w_j}{w_i} \right] \end{aligned}$$

(10)

Step 6: Consistency Ratio (CR) refers to the total consistency of the AHP calculation. If $CR > 0.1$, the importance of each criterion must be reconsidered (Saaty, 1980). The formula is given in Eq. 9:

$$\begin{aligned} CR = \frac{CI}{RI} \end{aligned}$$

(11)

where RI stands for Random Index, a value according to the matrix order.

3.5.2 TOPSIS

The TOPSIS method, introduced by Hwang and Yoon (1981), emphasizes the significance of calculating the distance of each alternative from the positive ideal solution and the negative ideal solution. The procedural steps for the TOPSIS method include:

Step 1: The evaluation matrix represents the alternatives and a set of criteria, where $(A_{ij})$ defines ratings.

$$\begin{aligned} \begin{array}{c|cccc} A & {\textrm{Criteria}}\, 1 & {\textrm{Criteria }}\,2 & \ldots & {\textrm{Criteria}}\, n \\ \hline {\textrm{Alternatives}}\, 1 & A_{11} & A_{12} & \ldots & A_{1 n} \\ {\textrm{Alternatives}}\, 2 & A_{21} & A_{22} & \ldots & A_{2 n} \\ \ldots & \ldots & \ldots & \ldots & \ldots \\ {\textrm{Alternatives}}\, m & A_{m 1} & A_{m 2} & \ldots & A_{m n} \end{array} \end{aligned}$$

Step 2: Calculate the normalized values $(R_{ij})$ using the alternatives m and criteria n by using Eq. 12

$$\begin{aligned} R_{ij} = \frac{A_{ij}}{\sqrt{\sum _{i=1}^m A_{ij}^2}} \quad i=1,2,3,\dots ,m; \, j=1,2,\dots ,n \end{aligned}$$

(12)

Step 3: Assign weights to the criteria and calculate the weighted normalized values $(V_{ij})$ by using Eq. 13

$$\begin{aligned} V_{ij} = W_j \cdot R_{ij}, \quad i=1,2,3,\ldots ,m; \, j=1,2,\ldots ,n \end{aligned}$$

(13)

where $W_j$ represents the weight of the $j^{th}$ criterion.

Step 4: Determine the best $(s^+)$ and the worst $(s^-)$ performance values for each criterion using Eq. 14.

$$\begin{aligned} \begin{aligned}&s^+ = \max _{i}(V_{ij}) \quad \text {for each criterion } j, \\&s^- = \min _{i}(V_{ij}) \quad \text {for each criterion } j. \end{aligned} \end{aligned}$$

(14)

Step 5: Calculate the distance of each alternative from the positive ideal solution $(D_i^+)$ and the negative ideal solution $(D_i^-)$ using Eq. 15.

$$\begin{aligned} \begin{aligned} D_i^+&= \sqrt{\sum _{j=1}^n (V_{ij} - s_j^+)^2}, \quad i=1,2,\ldots ,m, \\ D_i^-&= \sqrt{\sum _{j=1}^n (V_{ij} - s_j^-)^2}, \quad i=1,2,\ldots ,m. \end{aligned} \end{aligned}$$

(15)

Step 6: Calculate the relative closeness to the ideal solution $(C_i)$ for each alternative using Eq. 16.

$$\begin{aligned} C_i = \frac{D_i^-}{D_i^- + D_i^+}, \quad i=1,2,\ldots ,m, \quad 0 \le C_i \le 1. \end{aligned}$$

(16)

The alternative with the highest $C_i$ value is considered the best choice.

3.5.3 VIKOR

The Mardani et al. (2016) method is a MCDM technique designed to identify a compromise solution by ranking and selecting from a set of alternatives in the presence of conflicting criteria. The steps of the VIKOR Method is as follows,

Step 1: Determine the Best and Worst Values for Each Criterion Identify the best ($f_i^*$) and worst ($f_i^-$) values for each criterion among all alternatives using Eq. 17.

$$\begin{aligned} & Benefit \ criteria: f_i^* = \max _j f_{ij},\quad f_i^- = \min _j f_{ij}.\nonumber \\ & Cost \ criteria: f_i^* = \min _j f_{ij},\quad f_i^- = \max _j f_{ij}. \end{aligned}$$

(17)

Step 2: Calculate the $S_j$ and $R_j$ Values For each alternative, calculate $S_j$ and $R_j$ using the Eqs. 18 and 19:

$$\begin{aligned} S_j= & \sum _{i=1}^n \frac{w_i (f_i^* - f_{ij})}{f_i^* - f_i^-}.\end{aligned}$$

(18)

$$\begin{aligned} R_j= & \max _{i} \left[ \frac{w_i (f_i^* - f_{ij})}{f_i^* - f_i^-} \right] . \end{aligned}$$

(19)

where $w_i$ are the weights of the criteria.

Step 3: Identify the Best and Worst $S_j$ and $R_j$ Values Determine $S^*$, $S^-$, $R^*$, and $R^-$, by Eqs. 20 and 21:

$$\begin{aligned} & S^* = \min _j S_j,\quad S^- = \max _j S_j,\end{aligned}$$

(20)

$$\begin{aligned} & R^* = \min _j R_j,\quad R^- = \max _j R_j. \end{aligned}$$

(21)

Step 4: Compute the $Q_j$ Values for each alternative using Eq. 22.

$$\begin{aligned} Q_j = \frac{v(S_j - S^*)}{(S^- - S^*)} + \frac{(1-v)(R_j - R^*)}{(R^- - R^*)}, \quad j = 1,2, \ldots , J, \end{aligned}$$

(22)

where v is a weight representing the strategy of the majority of criteria (commonly $v = 0.5$).

Step 5: Rank the alternatives based on the values of $Q_j$, from the smallest to the largest. The alternative with the minimum $Q_j$ value is considered the best solution.

3.5.4 Grey theory

Grey Theory (Aruldoss et al. 2013) is widely used in systems analysis, decision-making, and forecasting, particularly when data is limited, incomplete, or uncertain. It focuses on the development of models that can handle such uncertainty by generating and exploiting the grey system’s inherent patterns. The steps of the Grey Theory are as follows:

Step 1: The process begins with the creation of a data set based on criteria $C_0 = \{C_1, C_2, C_3, \dots \}$. This set represents the different criteria considered in the analysis.

Step 2: Determination of Comparison Data For each criterion, comparison data $C_i = \{C_{i1}, C_{i2}, C_{i3}, \dots \}$ is determined, showcasing the performance values of each alternative against the criteria, where $i = 1,2,3, \dots , k$, and k defines the number of alternatives.

Step 3: Calculation of Performance Indicators, indicators are calculated to evaluate each criterion’s maximum, minimum, and optimum value performance:

Maximum performance indicator in Eq. 23

$$\begin{aligned} \begin{aligned} V_i(l) = \frac{V_i(l) - \min V_i(l)}{\max V_i(l) - \min V_i(l)} \end{aligned} \end{aligned}$$

(23)

Minimum performance indicator in Eq. 24

$$\begin{aligned} \begin{aligned} V_i(l) = \frac{\max V_i(l) - V_i(l)}{\max V_i(l) - \min V_i(l)} \end{aligned} \end{aligned}$$

(24)

Optimum value performance indicator in Eq. 24

$$\begin{aligned} \begin{aligned} V_i(l) = 1 - \frac{|V_i(l) - U_i|}{\max |V_i(l) - U_i|} \end{aligned} \end{aligned}$$

(25)

The normalized data is then calculated from these equations.

Step 4 involves the Calculation of Distance Between Data Sets. The distance between data sets is computed as $\Delta _i = (|d_{01} - d_{i1}|, |d_{02} - d_{i2}|, \dots , |d_{0m} - d_{im}|)$, where the global maximum $(\Delta _{\max })$ and the global minimum $(\Delta _{\min })$ are identified.

Step 5: Transformation into Grey Relational Coefficient, Each data point in the difference set is transformed into a Grey Relational Coefficient using Eq. 26:

$$\begin{aligned} C_i \gamma _i(j) = \frac{\Delta _{\min } + \xi \Delta _{\max }}{\Delta _i(j) + \xi \Delta _{\max }} \end{aligned}$$

(26)

here, $\Delta _{i(j)}$ is the $j^{th}$ value in the difference set, and $\xi$ (between 0 and 1) is used to diminish the effect of $\Delta _{\max }$, the extreme value in the data set, typically taken as 0.5.

Step 6: Calculation of Grey Relational Grade, The Grey Relational Grade of alternative i is calculated using Eq. 27:

$$\begin{aligned} r_i = \sum _{n=1}^m (\gamma _i(n) \cdot w(n)) \end{aligned}$$

(27)

where w(n) represents the weight of the $n^{th}$ criterion.

Step 7: Finally, the criteria are ranked according to their Grey Relational Grade. The priority ranking is obtained, and the best alternative is selected based on this ranking.

3.5.5 ELECTRE

The ELECTRE (Figueira et al. 2016) methods (I, II, and III) are part of a family of multi-criteria decision-making (MCDM) methods that differ primarily in how they handle decision-making criteria and alternatives. While ELECTRE I and II focus on the use of concordance and discordance matrices, ELECTRE III introduces the principle of fuzzy logic, utilizing preference and indifference thresholds to determine concordance and discordance indexes. The steps for The ELECTRE I Method are as follows:

Step 1: The criteria are coded on numerical scales with identical ranges to allow for comparison. This step is not always simple and may require a common scale to be constructed. The coding should justify the use of the max operator introduced later to model discordance.

Step 2: For each pair of actions (a, b) in the set A, where $a \ne b$, compute the concordance index using Eq. 28:

$$\begin{aligned} c(a S b)=\sum _{\left\{ j: g_j(a) \geqslant g_j(b)\right\} } w_j \end{aligned}$$

(28)

$\{j: g_j(a) \ge g_j(b)\}$ is the set of indices for all the criteria belonging to the concordant coalition with the outranking relation aSb. The concordance index represents the strength of the coalition supporting the assertion "a is at least as good as b".

Step 3: For each pair of actions (a, b) in the set A, where $a \ne b$, compute the discordance index using Eq. 29:

$$\begin{aligned} d(a S b)=\max _{\left\{ j: g_j(a)<g_j(b)\right\} }\left\{ g_j(b)-g_j(a)\right\} \end{aligned}$$

(29)

The discordance index measures the strength of the coalition against the assertion "a is at least as good as b".

Step 4: An action “a outranks b” (aSb) if and only if:

• 1. The concordance index c(aSb) is greater than or equal to a given concordance level s, where s is typically in the range $[0.5, 1-\min _{j \in J} w_j]$.

• 2. The discordance index d(aSb) is less than or equal to a given discordance level v.

Step 5: Exploitation of the outranking relation, Identify a subset of actions (kernel) from which the decision-maker can select the final action. Actions outside the kernel are outranked by at least one action in the kernel. If the graph contains no direct cycles, there exists a single kernel forming a partition on A. If the graph contains direct cycles, a preprocessing step is needed to reduce maximal direct cycles to singleton components, forming a partition on A. A new preference relation $\succ$ is defined on the partition $\bar{A} = \{\bar{A_1}, \bar{A_2}, \ldots \}$.

$$\begin{aligned} & \bar{A}_p \succ \bar{A}_q \Leftrightarrow \exists a \in \bar{A}_p \text{ and } \exists b \in \bar{A}_q \text{ such } \text{ that } \nonumber \\ & a S b \text{ for } \bar{A}_p \ne \bar{A}_q \end{aligned}$$

(30)

In ELECTRE I, actions forming a cycle are considered indifferent, which may be criticized. ELECTRE IS was designed to address this issue.

3.5.6 PROMETHEE

The PROMETHEE (Preference Ranking Organization Method for Enrichment Evaluations) method is a multi-criteria decision-making (MCDM) approach that was developed by Brans and Vincke (1985) (Brans and Vincke 1985). Different versions of PROMETHEE were introduced ( Brans & Mareschal, 1991 Brans and Mareschal 1992, 1994 Brans and Mareschal 1994, 1995 Brans and Mareschal 1995). PROMETHEE evaluates various alternatives based on multiple criteria. It is distinguished by its use of preference functions to translate deviations between alternatives into preference degrees, facilitating a comprehensive and nuanced comparison of alternatives. The steps of the PROMETHEE Method are as follows:

Step 1: Determination of Preference Functions, The process begins by determining a specific preference function $P_j(a, b)$ for each criterion j. This function translates the deviation between the evaluations of two alternatives a and b into a preference degree ranging from 0 to 1, based on the equation provided by Murat et al. (2015), the equation is as follows 31:

$$\begin{aligned} P_j(a, b) = G_j\{f_j(a) - f_j(b)\} \end{aligned}$$

(31)

where $f_j(a)$ and $f_j(b)$ represent the scores of alternatives a and b on criterion j, respectively.

Step 2: Assignment of Weights to Criteria, Relative importance or weights of each criterion are assigned to reflect their significance in the decision-making process.

Step 3: Computation of Overall Preference Index, An overall preference index $\pi (a, b)$ is computed by taking into account all criteria. This index is based on the Eqs. 32, 33, 34, 35:

$$\begin{aligned} & \begin{aligned} \pi (a, b)&= \sum _{j=1}^k W_j P_j(a, b) \\ \end{aligned}\end{aligned}$$

(32)

$$\begin{aligned} & \begin{aligned} \varphi ^{+}(a)&= \frac{1}{n-1} \sum _b \pi (a, b) \\ \end{aligned}\end{aligned}$$

(33)

$$\begin{aligned} & \begin{aligned} \varphi ^{-}(a)&= \frac{1}{n-1} \sum _b \pi (b, a) \\ \end{aligned}\end{aligned}$$

(34)

$$\begin{aligned} & \begin{aligned} \varphi (a)&= \varphi ^{+}(a) - \varphi ^{-}(a) \end{aligned} \end{aligned}$$

(35)

where $W_j$ represents the weight of criterion j, and n is the number of alternatives.

Step 4: involves the Calculation of Positive and Negative Preference Flows. The positive preference flow, denoted as $\varphi ^{+}(a)$ measures the extent to which an alternative a is globally preferred over all other alternatives. Conversely, the negative preference flow, $\varphi ^{-}(a)$ assesses the degree to which an alternative a is globally preferred by all other alternatives.

Step 5: involves determining Partial and Complete Rankings. The PROMETHEE I method achieves a partial ranking by intersecting two distinct rankings: the first is created by arranging actions based on their positive flow scores in a descending order, and the second is formed by organizing actions according to their negative flow scores in an ascending order. An optimal action is characterized by having a positive preference flow of 1 and a negative preference flow of 0.

Step 6: Selection of the Best Alternative, The best alternative is selected based on the comprehensive evaluation of positive and negative preference flows, taking into account the relative importance of each criterion.

3.6 Optimum road route generation

The majority of road generation methods are based on least-cost algorithms (Sarı and Sen 2017; Magara et al. 2021). Remote sensing along with least cost path analysis (LCPA) is a proven tool for optimal route selection (Larbi et al. 2018). Least-cost path analysis (LCPA) based on shortest path algorithms is usually employed to find optimal road route alignment in the raster data in the GIS environment (Douglas 1994) (Douglas 1994). LCPA provides routes according to the desired parameter-oriented approach (Sarı and Sen 2017). LCP algorithm is used to find the least cost between two locations, where the least cost path would be minimum distance or minimum construction cost depending on the user (Digra et al. 2022; Kang and Lee 2017).

A study conducted by E. K. Larbi et.al propose an optimum route of length 53.12 km has been selected which is more than 4.69 km of the existing road but the fastest route between Bonsa and Bogoso using APH and LCPA (Larbi et al. 2018) (Fig. 20).

Fig. 20

A case study area with LCP result and buffered search corridor. (Nadir 2021)

A study conducted by W. H. Nadir integrates the GIS-GA (Geographic Information System-Genetic Algorithm) model to locate and configure a corridor-based optimum horizontal highway alignment (Nadir 2021). Manoj K. Jha et al. also utilize GIS, genetic algorithms (GAs) along with visualization techniques to optimize highway development of Maryland on the basis of minimization of highway cost (Jha et al. 2001). M. Rybansky proposed an approach that combines Delaunay triangulation, Voronoi graphs, and Dijkstra’s algorithm to determine the quickest or shortest paths, considering factors like tree stand structure, vehicle dimensions, and terrain constraints to effectively map out optimal routes for emergency scenarios (Rybansky 2014). study by Manoj K. Jha et al., adapts previous highway optimization strategies for rail transit, utilizing GIS for evaluating costs and environmental impacts (Jha et al. 2007).

Figure 21 shows the popular optimization algorithms used throughout the literature. among these algorithm’s least cost path algorithm is the most popular one but there is no clear explanation of the process and outline of the algorithm is given, also many research papers use this term to describe algorithms like Dijkstra (Jha et al. 2007) and A* (Dijkstra 2022).

Fig. 21

Optimization algorithm paper wise count

According to the research conducted by Ju Young Kang et al. Dijkstra algorithm is more useful when we are dealing with the GIS raster maps as compared to the A* algorithm, as A* can not guarantee finding the shortest path. Unlike video games or autonomous vehicles, real-time search is not important while optimum solution is a key requirement (Douglas 1994). However, according to research conducted by Marcelo Otone Aguiar et al., D’Esopo-Pape is more efficient than algorithms like Dijkstra, and Bellman-Ford (Aguiar et al. 2021).

Following are some of the most popular cost optimization algorithms according to the literature:

Algorithm 1

D’Esopo-Pape algorithm jakobkogler et al. (2022)

Algorithm 2

A* algorithm Candra et al. (2021)

Algorithm 3

Bellman-Ford algorithm Patel and Bagar (2014)

Algorithm 4

Dijkstra’s algorithm Patel and Bagar (2014)

Algorithm 5

Genetic algorithm (GA) https://www.cs.ucc.ie/~dgb/courses/tai/notes/handout12.pdf

4 Discussion

This paper provides unbiased and critical reviews of the literature. The research utilizes the PRISMA method which is highly respected for its systematic and structured approach which enhances the credibility and trustworthiness of the review. The use of multiple major databases and the ’backward snowballing’ technique increases the chance of finding the most relevant research available. Inclusion and Exclusion Criteria Clearly defined filters ensure including research that genuinely aligns with research aims, maintaining the focus defined in Table 3.

The research questions mentioned in Table 1 are crisp, unambiguous and unique which are not being considered in any other research paper and these questions are essential to understand road route optimization, deep learning techniques for LULC and application of MCDM techniques.

The analysis of publication trends highlights intriguing dynamics within the field of road route alignment optimization research. The steady increase in publications on road route alignment, with a dip in 2022, suggests a maintained interest in this domain. Notably, the growth of publications related to LULC using deep learning in 2023 reveals a promising trajectory for incorporating advanced computational techniques also there is an influence of increasing computational power, making DL more feasible. This surge could be due to the increased accuracy and efficiency that deep learning offers in analyzing complex land cover datasets. Conversely, the decline in articles focused on MCDM techniques from 2020 onwards warrants further investigation. It’s possible that the COVID-19 pandemic’s disruptions to fieldwork and data collection may have contributed to this decrease. Additionally, researchers might be re-evaluating MCDM techniques in light of new challenges posed by the pandemic, such as the rise of e-commerce and potential changes in traffic patterns. Alternatively, there could be a shift towards integrating MCDM with deep learning methods or exploring other optimization techniques altogether which are better suited to the increased computational resources now available.

The higher number of articles originating from India, China, Turkey, and the USA reflects a correlation between infrastructure development priorities and research output. This aligns with the importance of road networks as a core component of transportation systems, which are vital for economic growth, particularly in developing and developed nations.

The presented generalized framework offers a solid basis for effective road route planning and optimization by combining deep learning and MCDM techniques. Its comprehensive and adaptable structure suits diverse contexts and project needs. The iterative weight assignment within the MCDM phase allows the exploration of different scenarios and prioritization of various criteria, leading to informed decision-making. The emphasis on deep learning for land use classification and GIS-based spatial analysis highlights the framework’s capability for accurate and efficient route optimization. However, it’s important to be aware that the framework’s success depends on the quality and lack of bias within input datasets. Computational intensity should be a consideration. Overall, this generalized framework serves as a valuable starting point, with the potential for continuous refinement to address the challenges of sustainable transportation and infrastructure development.

The outlined sustainability criteria emphasize the multifaceted nature of sustainable road construction. Prioritizing construction on stable, less sensitive lands underscores the need to rehabilitate existing infrastructure before expanding into pristine ecosystems, thus minimizing habitat destruction. Life-cycle assessments (LCA) may help in informed decision-making about materials and techniques that minimize environmental impacts. Community engagement plays a vital role; incorporating community voices in the planning process helps mitigate social disruptions and ensures projects address local concerns. Government agencies face the complex task of balancing these sustainability considerations with the need for transportation infrastructure development. Finding this balance often necessitates careful assessments of the economic benefits of new roads against their potential environmental and social costs. By considering sustainability in all phases of road construction, we can create a transportation system that minimizes environmental harm, supports economic health, and prioritizes the well-being of communities.

The field of Land Use/Land Cover (LULC) classification has seen an evolution in techniques. Initially, traditional methods like Maximum Likelihood, Support Vector Machines, and Decision Trees were used, but these face limitations when handling large datasets or complex patterns. Deep Learning (DL) has come to dominate the field due to its ability to process vast amounts of data and uncover intricate patterns, which translates into higher accuracy in classifying land features. Techniques like Deep Belief Networks, Generative Adversarial, Stacked Autoencoders, Recurrent Neural Networks, and Convolutional Neural Networks are all being explored for LULC applications. The quality of the dataset plays a significant role in DL success; this means selecting appropriate data sources and rigorously prepping input through noise reduction and preprocessing. LULC classification finds valuable applications across sectors. It can be used in urban planning, resource management, and sustainable development but also extends to fields like oil and gas pipeline routing. LULC analysis can serve as a crucial backbone for many linear engineering applications, helping classify land cover types to support informed decision-making about infrastructure development and minimize environmental impact.

The integration of MCDM techniques with GIS provides a powerful method for selecting optimal road and highway alignments. These methods offer a structured approach to handle complexities, carefully balancing factors like environmental impact, economic cost, social concerns, and technical feasibility. MCDM is ideal for this application as it allows for the consideration of multiple, often conflicting criteria, incorporates expert judgment and stakeholder input, and increases transparency in the decision-making process. Selecting the most relevant criteria, assigning appropriate weights, and choosing the right MCDM techniques (such as TOPSIS, AHP, VIKOR, ELECTRE, Grey Theory, and PROMETHEE) are crucial for successful outcomes. By carefully utilizing these strategies, MCDM and GIS enable well-informed and sustainable road alignment choices.

Choosing the most suitable algorithm for optimal road generation depends on several factors, including the prioritization for the road (shortest distance, minimizing construction costs, environmental impact reduction, etc.), the data format available (grid-based maps, linear maps, etc.), and the trade-off between finding the absolute best solution and obtaining a solution quickly. For guaranteeing the absolute shortest path, Dijkstra’s algorithm is the reliable choice, especially when negative edge weights aren’t present. A* can be faster, particularly with a well-designed heuristic, and also handles negative edge weights. When road network planning problems become complex for traditional methods to efficiently navigate vast solution spaces, genetic algorithms provide a powerful alternative. These nature-inspired algorithms excel at searching these complex landscapes and identifying high-quality solutions. The D’Esopo-Pape algorithm often proves very efficient for road generation and addresses negative edge weights; however, it’s crucial to be aware of its potential for exponential time complexity in rare worst-case graph structures. The ideal choice ultimately depends on the problem’s constraints, the type of optimization desired, the data format, and the relative importance of guaranteed optimality versus computational speed. Investigating hybrid approaches and emerging optimization techniques could further improve road generation solutions in the future.

5 Future scope and research directions

The findings of this SLR illuminate several pathways for future research and development in the field of road route optimization using GIS, deep learning, and MCDM techniques. These pathways not only address the existing gaps identified in the current body of literature but also aim to leverage the rapid advancements in technology and computational methods. The future scope of this research domain can be outlined as follows:

• 1. Integration of Emerging Technologies: Future studies should explore the integration of newer geospatial technologies and emerging deep learning models into the road route optimization process. This includes the application of cutting-edge AI techniques such as Generative Adversarial Networks (GANs) and Reinforcement Learning in the context of GIS for more accurate and dynamic route planning, in domains like agriculture (Vafaeinejad et al. 2025) and satellite-captured images (Sharifi and Safari 2025).

• 2. Enhanced MCDM Approaches: There is a clear need for the development of more sophisticated MCDM frameworks that can handle the complexities and multi-faceted nature of road route optimization. Future research should focus on creating more nuanced MCDM models that incorporate a wider range of criteria, including climate change impacts, social equity, and future urban growth patterns.

• 3. Cross-Disciplinary Approaches: Combine insights from urban planning, environmental science, social sciences, and computer science can lead to more holistic and sustainable road route optimization methodologies. This could involve developing new collaborative frameworks that allow for the incorporation of diverse perspectives and expertise.

• 4. Data Availability and Quality: There remains a significant challenge in the availability and quality of geospatial data required for road route optimization. Future research should address these challenges by developing new methods for data collection, processing, and sharing, particularly in regions where data are scarce or of poor quality.

• 5. Case Studies and Real-world Applications: To validate and refine the proposed methodologies, there is a need for more comprehensive case studies and real-world applications. Future studies should apply the developed frameworks to actual road planning projects, providing insights into their practicality, effectiveness, and areas for improvement.

• 6. Sustainability and Resilience: Future research should place a greater emphasis on sustainability and resilience in road route planning. This includes studying the long-term impacts of road infrastructure on ecosystems, communities, and climate change, and developing routes that are not only optimized for current conditions but also adaptable to future changes.

• 7. Global and Regional Studies: Given the varying geographical, environmental, and socio-economic conditions across the globe, there is a need for more studies that focus on specific regions or compare different geographical contexts. This will help to develop more localized solutions and understand how global trends are manifested at the regional level.

By addressing these areas, future research can significantly contribute to the development of more efficient, sustainable, and equitable road route optimization strategies, ultimately leading to better infrastructure planning and development worldwide. While this study provides a comprehensive synthesis of existing literature, it is not without limitations. As a SLR, this research does not propose or empirically validate a new algorithm or model. Instead, it focuses on identifying trends, gaps, and opportunities within the domain. Consequently, experimental validation and performance benchmarking of specific techniques are beyond the scope of this work. Additionally, integrating deep learning and MCDM methods within GIS-based frameworks often entails significant computational overhead. Limitations such as the dependency on commercial tools for LULC classification, lack of open-source comparability, and challenges related to scalability and processing efficiency in real-world applications were observed. Future research should seek to overcome these constraints by developing open-access tools, optimizing algorithms for performance, and exploring scalable solutions suitable for diverse and large-scale infrastructure projects.

6 Conclusion

This systematic literature review provides a comprehensive synthesis of research focused on optimizing road route alignment. It highlights the powerful potential of integrating Geographic Information Systems (GIS), deep learning, and Multi-Criteria Decision Making (MCDM) techniques for tackling complex challenges associated with sustainable road design development. The findings emphasize the importance of considering environmental, social, economic, and engineering factors to align road projects with broader sustainable development goals.

Our in-depth examination of 157 papers using the PRISMA methodology reveals that while researchers have extensively explored various algorithms and performance metrics, a truly end-to-end framework is still needed. We identified a reliance on commercial software for tasks like Land Use and Land Cover (LULC) classification as a potential barrier to comparative analysis. Additionally, the lack of specificity regarding optimization algorithms used in commercial software hinders a clear assessment of optimization levels.

Despite its comprehensive and systematic approach, this research faces inherent computational limitations associated with integrating deep learning techniques and multi-criteria analysis into GIS-based frameworks. The computational intensity, influenced by data quality, algorithmic complexity, and resource availability, poses practical challenges for large-scale real-world applications. Future studies should therefore prioritize enhancing computational scalability and processing efficiency, potentially through the integration of more advanced computational methodologies and optimization algorithms. Addressing these limitations will not only enhance the robustness of the framework but also significantly broaden its applicability across diverse geographical and infrastructural contexts.

To address these issues, we propose a generalized framework designed to streamline road route optimization. The framework outlines a systematic approach encompassing problem definition, criteria selection, data preparation, deep learning-based LULC classification, MCDM, and least-cost path analysis. This structured framework serves as a foundation for future research and real-world implementation.

This study makes several key contributions to the field. Firstly, by thoroughly analyzing existing literature, it offers a comprehensive, statistically-backed overview of the current state of the field. This includes revealing trends in research focus, geographical distribution of studies, and the evolution of methodologies. Secondly, the study demonstrates the efficiency of deep learning in significantly improving the accuracy and efficiency of LULC classification, a crucial step in identifying and evaluating potential road routes. Thirdly, it showcases the increasing adoption of MCDM techniques for a more holistic approach to road alignment, incorporating diverse environmental, economic, and social factors. Lastly, by highlighting the potential for further integration of deep learning and GIS methods, the study pinpoints an exciting, underexplored avenue within the field.