Review of road selection methods for the purpose of multiscale mapping

These methods use the importance of roads based on attribute information, such as road class and road type, road category or any other attribute which can hierarchize this thematic layer. The semantic information of roads is often employed as auxiliary features in conjunction with geometric and topological road features for road network selection (Tang et al, 2024). The implementation of this method relies only on what the map designer needs to achieve.

The road selection process usually utilizes input data from the vector database. Selection is based on attributes and their values. An example of the semantic-based generalization was presented by Sielicka and Karsznia (2019) where General Geographic Object Database (GGOD), which includes detailed attributes of roads, such as class, number of lanes, management category, existence status, and surface material was used as a source database. The methodology applied a combination of attribute-based selection and network connectivity analysis to identify roads to be retained for small-scale maps. Higher-class roads, such as highways and expressways, were prioritized based on their attributes, while lower-class roads were included only if they ensured overall network connectivity, such as linking to settlements or forming junctions with higher-class roads. The selection process was implemented in ArcGIS using various tools and algorithms to analyze connectivity as well as to evaluate road importance. The output was a simplified and symbolized road network that retained essential connectivity and functionality, designed for scales of 1:500,000 and 1:1,000,000. The results were evaluated according to the visual and functional consistency, ensuring the generalized road network aligns with the requirements of small-scale cartographic representation (Sielicka & Karsznia 2019).

The line type approach (more commonly known as stroke-based) assumes the connection between individual road segments using the conceptual grouping principle of good continuation (Thomson & Brooks 2000). At each intersection, a decision is made regarding which segment to connect to the next one. This decision is typically based on the deflection angle between the considered segments (Weiss & Weibel 2014). Generally, a smaller angle is preferred; however, this is not always the case. The attributes stored in the database, such as road class or road name, can supplement this decision-making process. The creation of strokes allows the combination of the road segments into longer sections, which are then selected to be displayed on the map, while shorter segments are omitted.

Due to significant differences in the principles of generalization between rural and urban road networks, different methods are often employed for each. Such an approach was adopted by Touya (2010). In rural areas, the author relied on enriched data describing road strokes when deciding on road selection. The preliminary processes such as data enrichment (Brassel & Weibel 1988) and the proper detection of complex junctions, highway interchanges, and crossroads were crucial for appropriate stroke creation. The selection was based on the shortest paths between attraction points (e.g., residential areas to commercial areas). Attraction points came from facility data, weighted by the importance. If facility points are not available, random points can be used (Ruas & Morisset 1997). Travel time estimation, not just road length, was used for path computation. Factors such as road class, sinuosity, and town crossing influenced travel time. Roads were selected based on road use estimation, stroke length, T-node crossings, and dead-end characteristics.

Another interesting example of stroke type data application for the purpose of map generalization was proposed by Yang et al. (2011). The authors presented a method of generating hierarchical strokes in urban street networks to ensure good continuity. The method addressed such challenges as detecting and tracking dual carriageways, especially when auxiliary streets connect them, and identifying complex street junctions to maintain stroke continuity during stroke generation. Then, the network was used as a source for stroke ranking. Strokes were ranked based on length and centrality measures. These measures were integrated into a comprehensive indicator for hierarchical selection. Last but not the least, the task was to maintain the connectivity of the source urban street network. The minimum spanning tree (MST) method was thus employed to connect disjointed strokes and regain overall road network connectivity.

Further methodological solution, proposed by Kim et al. (2019), focused on on-demand road generalization to create sketch maps. This consisted of four main steps: (1) identifying input facilities, selected by the user, to be surrounded by a road network; (2) creating suitable road networks; (3) rearranging selected objects; and (4) generating a final map. The emphasis was on road generalization, specifically on generating global and local stroke networks structures. The system took the strokes from a database within a given map area and created a global stroke network by selecting suitable strokes based on predetermined conditions. This network partitioned the map into sub-areas, within which a local stroke network was created for areas containing the facility. Both stroke networks were represented using GeoJSON format; the resulting map with road generalization was illustrated using OpenStreetMap and displayed using Leaflet. The algorithm selected the minimum number of necessary strokes from the local stroke network based on the reachability to facilities and average angles with other selected strokes, ensuring high reachability and avoiding redundant detours.

Liu and Li (2019) proposed an algorithm for stroke generation in road networks that incorporates geometric and structural properties using a feature-based information entropy model. The method addresses limitations of traditional stroke generation techniques, which rely on fixed angle thresholds and disregard road network diversity and structural variability. The proposed approach used an entropy-based model to calculate road network information at the road segment and neighborhood levels, reflecting geometric characteristics and topological connections under varying angle thresholds. The algorithm thus improved stroke generation technique by introducing steps to compute connectivity, identify merging candidates based on deflection thresholds, and optimize computational efficiency. Although it still employs deflection angles to merge road segments into strokes. Furthermore, the Douglas-Peucker algorithm was used to simplify the information volume curve, enabling the determination of optimal angle thresholds for stroke generation across different road network morphologies. The authors determined optimal angle threshold ranges for stroke generation in each test network of different morphology (e.g., 45°–48° for Monaco, 54°–63° for Chicago, and 61°–63° for Moscow). Results demonstrated that these thresholds led to improved connectivity and reduced suspension of strokes compared to fixed-threshold methods. Authors highlighted the importance of adapting thresholds to specific road network structures for effective stroke generation.

An interesting and recently used road selection approach is an analytical hierarchy process (AHP). AHP constitutes one of the multicriteria decision-making methods. It helps to find optimal decisions by considering multiple values as well as taking into consideration the user preferences. This approach was recently used for road selection by Han et al. (2020). The authors used points of interest (POIs) to build indicators of contextual characteristics and calculate stroke importance with the use of the AHP model. Roads were selected according to the stroke importance, as well as on further criteria of density and overall road network connectivity maintenance. The scope of this research covered road selection from 1:5 000 to 1:200 000 and based on the resulting map the authors concluded that using the AHP method helps to preserve the structure and characteristics of the source road network.

The mesh- or block-based approach indirectly enables to select the most relevant roads by consolidating less significant road segments within a specified area. In this method, adjacent road segments form a mesh cell or block. The retained road count is determined by a density threshold, which can be either fixed or dynamic, based on neighboring mesh information. One of the first examples of this approach was proposed by Chen et al. (2009). Their approach involved an iterative process, in which the mesh with the highest density was identified, and its least important road segment was removed. The remaining mesh parts are then merged with the closest mesh to form a larger one. The elimination of road segments follows a priority order based on parameters such as road class, length, or stroke degree.

A similar block-based approach was suggested by Gülgen and Gökgöz (2011). In this method, only roads from the lower classes (less significant than highways, for example) are considered. Blocks are created in a cycle involving two stages (Gülgen & Gökgöz 2011). Here, the primary objective is to maintain the space allocated for individual buildings, which varies based on the target scale. Blocks are then categorized based on “minimum block space” and “minimum block area.” Problematic blocks undergo generalization, starting from the smallest, using algorithms such as the smallest neighbor block algorithm (which merges the two smallest blocks) and the minimum-weighted common boundary algorithm (which removes the least important boundary between blocks).

Recently, some new enhancements to the original mesh method have been introduced. Initially, the starting mesh is selected based on its density and boundary characteristics. Further, instead of a single mesh, a pair of meshes with the highest density is considered as the starting mesh (Wu et al. 2022). Each iteration of mesh elimination then follows the stroke connection order rather than descending mesh density. This iterative method processes all small meshes within the aggregated areas until the densities of the newly formed meshes exceed the assumed threshold (Wu et al. 2022).

Graph theory can be applied to road networks by initially converting the network into a graph structure, which is a straightforward task: road segments connecting intersections become edges, while intersections represent nodes. Graphs can articulate topological relationships between objects, allowing for calculations of measurements such as shortest path, betweenness, closeness, degree and minimum spanning tree. These calculations can be later used as a importance metrics for road selection (Mackaness & Beard 1993; Jiang & Claramunt 2004). Such workflow was followed by Jiang and Claramunt (2004). The authors proposed the methodology of generalizing urban street networks based on structural properties. This work aimed to simplify representations while preserving the fundamental characteristics of the network. The input to this approach was a detailed representation of the street network where vertices represented named streets and edges represented street intersections (Jiang & Claramunt 2004), encompassing geometric, topological, and semantic properties of streets. Streets were then ranked based on their structural importance using measures like connectivity, centrality, and hierarchy. The generalization process iteratively removed less significant streets while maintaining network connectivity and topological integrity. The output was a generalized network suitable for various scales, retaining the core structure and functionality of the urban street network.

Another approach which utilized graph structure for road selection was proposed by Chmiani et al. (2014). This method focused on generating consistent and connected representations of road networks across multiple scales for continuous zoomable maps. Road networks were represented as weighted graphs, with edge weights reflecting road importance based on attributes like road class or centrality measures. The methodology revolved around formulating the problem as an optimization task, aiming to identify an optimal sequence of edge removals that maximizes the aggregated weights of retained edges across all scales while maintaining network connectivity at every step. The input constituted a connected weighted graph that represented a road network at the largest scale. The output was a sequence of road network generalizations that progressively simplified the network while preserving its overall structure and connectivity.

Lyu et al. (2022) proposed a slightly different method, which aimed at generalizing road networks based on residential areas to maintain spatial distribution consistency and spatial relations between roads and residential areas. The proposed method involved six steps: clustering of residential areas, classification of roads and settlements, evaluating settlement importance, building a geographic network, removing redundant paths, and ensuring visual and topological connectivity. Settlement importance was determined based on area, residential density, Voronoi area, and road index. Direct and indirect paths between the most important settlements were identified using Dijkstra’s least-cost algorithm based on travel time. The final stage involved removing redundant roads either by simplifying adjacent settlement pairs or by evaluating and potentially replacing paths between adjacent settlements based on their costs and connectivity.

These methods incorporate the artificial intelligence elements to make the selection process more efficient. The implementation of advanced methods can be conducted in various ways. From self-organizing NN (Jiang & Harrie 2004) through supervised ML models such as DT, RF, SVM and finally the use of the most advanced DL models. A compelling example of ML utilization is the road selection research proposed by Karsznia et al. (2022), where models are compared with the official regulations of road selection provided by the Polish National Mapping Agency. The traditional approach, involving simple attribute selection, was compared with a novel approach that utilizes data enrichment and ML models for binary classification of road segments. The training data were collected with the assistance of experienced cartographers. The study employed DT, decision trees optimized with genetic algorithms (DTGA) and RF models.

Zheng et al. (2021) employed GCNs for road network selection, abstracting the road network into a graph structure. The dual graph representation of the road network was adopted, switching the roles of nodes and edges to better capture the network’s topology. This representation aligns with the tendency of GCNs to aggregate node characteristics. The authors considered only three attributes concerning roads: road type, length, and coordinates. The model structure comprised three main components: a graph-convolution block for extracting local spatial features, a fusion block for integrating global and local features, and a prediction block for forecasting the selected roads. Multiple model architectures were tested, including JK-Nets, ResNet, and DenseNet.

All methods mentioned in figure 2 and described in previous sections can be combined to exploit the advantages of various solutions. Examples of such applications are briefly explained in the following sections.

This involves the combination of the stroke or graph approach with meshes. By combining linear and areal structures, it is possible to keep both long road segments and road network density. Stroke generation allows for the creation of real-world long road segments, while graphs can supplement the information about networks using centrality measures. The mesh approach, on the other hand, allows for the tuning of the desired network density. Such a combination was proposed by Li and Zhou (2012), where two types of road network are formed. Meshes, strokes, and bridges between them were identified. In the first step, the areal hierarchies were traversed until the desired number of areal segments was selected or omitted (Weiss & Weibel 2014). In the second step, the linear segments were traversed in a top-down manner. When constructing linear networks, the search for the shortest path linking them to areal or linear networks was carried out.

To improve this novel method, Benz and Weibel (2014) decided to change the approach for linear transversion to bottom-up. They proposed a set of constraints that helped the final selection be more cartographically appropriate. The constraints indicated the fraction of source road segments to be retained (70% of source), which segments must be retained (most important ones, such as highways), and that the roads must not be disconnected locally. The stage of long stroke reconnection, roundabout recognition, and POI accessibility was introduced to keep the road network connected. To keep the road network density in urban areas higher than in rural areas, meshes were formed only in areas labeled as settlement areas.

Another interesting solution was presented by Zhang et al. (2017), where the authors proposed an improvement to the hybrid method, combining linear and areal representation. The approach enhances the traditional stroke generation algorithm by introducing stroke evaluation indicators and using a weighted Voronoi diagram for global control of road selection. First, the stroke generation algorithm was applied in a linear representation mode by using an ordinary least squares (OLS) model to consider overall information for the roads to be connected. Simplified indicators, such as road length and the number of connected roads, were then used to calculate road segment importance. Second, in the areal representation mode, the proposed method partitioned road networks and calculated road density based on weighted Voronoi diagrams. Three parameters – stroke importance (SI), stroke density threshold (DS), and selected network length (LS) – were considered for road selection. Strokes were generated and sorted based on their importance. A weighted Voronoi diagram was used to partition the road network and calculate density. Finally roads were selected based on stroke importance and density threshold, ensuring a reasonable spatial structure.

A combination of stroke and graph allows for the creation of long, connected road segments that resemble those appearing on paper maps. The graph approach enables measures to be used such as the “shortest path between random points,” or centrality measures such as betweenness, closeness, and degree. Additional information about the geometric characteristics of roads can enhance the selection process.

A prime example of incorporating the line approach with additional information, gathered from stroke and graph road structure, was presented by Weiss and Weibel (2014). The authors improved the selection algorithm based on betweenness centrality. The main contribution of this paper consisted of proposing four extensions that addressed the deficiencies of the basic centrality-based algorithm and led to a significant improvement of the results. The first extension ensured that roundabouts were properly detected and collapsed. The second added more semantic information to the stroke generation process, allowing longer and more plausible strokes to be built (Weiss & Weibel 2014). The third extension, using density-based clustering, detected areas of higher road density. As a result, the local threshold in these areas have been increased, leading to more road segments being eliminated from urban areas. The final extension addresses inappropriate dead ends by reconnecting roads that should be part of the network but were not included after the selection process. The first two extensions greatly improved the stroke creation process, while the third and fourth were implemented to maintain road density at a reasonable level and to avoid discontinuities in the road network.

One of the first methods which used such a combination was a GCN framework combined with mesh-line structure unit (MLSU) data structure (Xiao et al. 2024). The MLSU integrated the characteristics of both mesh and line structures, capturing road and polygonal mesh features within a unified framework. By combining roads and their adjacent meshes into MLSUs, the method facilitated the comprehensive feature extraction for both interconnected road networks and isolated roads. The implementation uses topology adaptive graph convolutional networks (TAGCN) to process the complex relationships within the MLSUs. TAGCN enhances the ability to extract higher-order neighborhood information, enabling the model to learn spatial and topological generalization rules directly from the data. The model was trained on a dataset of roads from the southern area of the United States, mapped at scales of 1:1,000,000 and 1:2,000,000, with MLSUs characterized by road-specific, mesh-specific, and interaction-based features. These features included road hierarchy, mesh density, geometric properties, and relationships between roads and meshes. The adoption of GCN reduced the reliance on manually defined rules and thresholds, allowing the system to derive generalization decisions autonomously.

A study of the structure of a benchmark proposed by Zhou and Li (2017) lets us see how capable of correct road selection the different supervised learning algorithms are. In this research, digital road networks from three study areas were considered. The source scales were 1:20 000 and 1:50 000, whereas the target scales for selection and benchmark were 1:50 000, 1:100 000, 1:200 000, and 1:250 000. The authors decided to compare nine different algorithms – namely: ID3, C4.5, CRT, RF, SVM, naive bayes (NB), k-nearest neighbor (KNN), multilayer perception (MP), and binary logistic regression (BLR). Information about the strokes was encoded in the following attributes: length, degree, closeness, betweenness, road class, and number of lanes. However, it is important to note that not all databases contained complete information; for example, some were missing the road class attribute. The output of the applied algorithm was the binary classification of roads. Overall, the experiment aimed to evaluate the effectiveness of various supervised learning approaches regarding the selective omission of road segments based on different input attributes and output classes.

Visual examination of the results is essential in research on cartographic generalization. The perceived correctness of outputs often relies on the expertise and subjective judgment of experienced cartographers. In the majority of the referenced works, the authors examined visual results of automatic selection based on their own judgements. Although there were cases where researchers enrolled other cartographers to consult or assess the quality of the results (Benz & Weibel 2014; Weiss & Weibel 2014; Sielicka & Karsznia 2019; Han et al. 2020). At times, the differences among various approaches are so subtle that they cannot be identified through visual inspection alone (Karsznia et al, 2024b). Although numerous methods of effective road thinning networks exist, the primary roads are generally retained, while the inclusion or exclusion of minor roads varies widely between the screened approaches. This situation also arises in manual generalization, where a map designer decides which roads to preserve based on the surrounding area characteristics, the presence of important nearby features, and personal expertise. To ensure consistent evaluation and facilitate direct comparison, most studies rely on a single target map as a standardized reference. Adopting a common dataset allows for systematic assessments of different methods under the same conditions, thereby supporting more objective and comparative analyses. This brings us to the most commonly used metrics, which align with adopting a single baseline for selection assessment to maintain simplicity.

Among the variety of metrics reported in the literature as F1 score, maximum similarity, mean improvement over baseline approaches, correlation coefficients, shape similarity, and others, the accuracy emerges as the most consistently and frequently employed performance indicator. Approximately 35% of the referenced works contained in appendix 1 point out the accuracy as the primary or “major metric,” making it a base for benchmarking performance across different approaches and scale transitions. The popularity of this metric can be attributed to its ease of application and straightforward interpretation. Accuracy reflects how often a particular method successfully selects or retains the “correct” features (e.g., roads, strokes, or mesh elements) relative to a ground truth or a reference dataset or map. In a field where the output (a generalized road network) needs to closely resemble an accepted standard, accuracy provides a clear, binary measure of right vs. wrong selections.

The second and yet most frequent metric appears to be F1 score and maximum similarity. F1 score combines the concepts of precision and recall into a single number, making it especially useful when one needs a balance between these two measures. Apart from accuracy, precision and recall are the basic metrics calculated for automatic classification tasks. As the F1 score constitutes the harmonic mean of precision and recall, it penalizes extreme differences between them. If precision is very high but recall is low (or vice versa), the F1 score will be lower. This ensures that the model is performing well both in detecting as many positive cases as possible (high recall) and ensuring that its positive predictions are correct (high precision). Other measures, like common stroke ratio, matching ratio, number of road segments, selected road length change, mean similarity with the target, road length overlap, average number of identical strokes, and mean consistency with the existing map, all yield percentage-based values derived from comparing selected road segments against a reference. Although these metrics emphasize slightly different aspects, such as the proportion of shared strokes, the extent of length overlap, or the consistency of selected features, they are fundamentally similar as they quantify how closely the resulting generalization aligns with a known standard, which can be ground truths defined by authors or target map that they used in the research.

During the earliest works, traditional ML classifiers such as SVM, DT, RF (Zhou & Li 2017; Karsznia et al. 2022; Karsznia et al. 2024a), back-propagation neural networks (BPNN) (Zhou & Li 2014), and logistic regression (Park & Huh 2019) were employed to guide the selection of features for map generalization. Within these frameworks, accuracy emerged as a major measure to determine how closely the generated output approximated a trusted reference dataset. These initial studies often relied on well-defined training samples and labeled data, enabling accuracy to serve as a simple and efficient indicator of the proportion of correctly retained or omitted objects. As the research concerning GeoAI and ML evolved, the attention shifted toward more complex, graph-based, and network related methods. GCN, graph attention networks (GAT), graphSAGE (Zheng et al. 2021; Xiao et al. 2024; Tang et al. 2024), and other hierarchical frameworks appeared, reflecting a growing recognition of the importance of topological relationships in generalization process. Although these models introduced greater sophistication, complex network structures and advanced feature embeddings, they continued to report accuracy as a core measure of quality.

In parallel, integrated neural architectures and approaches enriched with semantic features were further developing. In these approaches, the inclusion of domain-specific heuristics, semantic constraints, and specialized indicators such as satisfaction of hard constraints or maximum similarity scores offered more nuanced insights into a model’s performance.

Within the reviewed studies and proposed methods the major purpose of the generalization and specifically road selection is identified as “database generalization”, accounting for 64% of the studies reviewed. This category, spanning from the earliest contributions (e.g., Liu et al. 2010; Touya 2010) to the most recent investigations (e.g., Xiao et al. 2024; Tang et al. 2024), consistently highlights the emphasis on producing generalized, datasets capable of supporting a wide range of subsequent applications.

By focusing on database-oriented outcomes, these methods are intended for scalability and reproducibility, ensuring that the designed maps can be readily integrated into mapping workflows regardless of the map use. In contrast, other purposes appear less frequently and exhibit a more specialized focus. Topographic maps, for instance, comprise 12% of the studies, reflecting a dedicated interest in maintaining thematic integrity and spatial accuracy within distinct geographical contexts. “Map updates” and “navigation,” each accounting for 8%, point to more targeted or dynamic needs, such as ensuring temporal relevance or route optimization. Meanwhile, “general geographic map” and “urban road generalization,” each at 4%, highlight niche demands arising from particular environments or user scenarios (appendix 1). This distribution of thematic purposes suggests that while specialized interests occasionally shape methodological decisions, they do not overshadow the prevalent goal of database generalization.

In some studies, the improvements across multiple areas were presented (Weiss & Weibel 2014). In such cases, the accuracy values were included in appendix 1. Other measures indicate how well the described methods function as improvements over previously implemented ones (Wu et al. 2022; Yang et al. 2011; Touya 2010; Chen et al. 2009), how accurately they mimic existing map products or manual selections (Gülgen & Gökgöz 2011), or if the proposed improvements satisfy the specified constraints (Benz & Weibel 2014), making them more comparable to the rest of the metrics gathered in appendix 1. The comparison of different road network generalization approaches reveals their varying effectiveness as well as their applicability to multiple scales. While many publications tackled the problem of selection for 1:200 000 target scale (Li et al. 2012; Zhou & Li 2014; Weiss & Weibel 2014; Zhou & Li 2017; Han et al. 2020; Tang et al. 2024; Xiao et al. 2024) there are only a few publications which focus on small-scales (Karsznia et al. 2022; Karsznia et al. 2024a; Zheng et al. 2024). ML-based approaches, such as those presented by Karsznia et al. (2024) and Guo et al. (2023), generally demonstrate high accuracy across various scales. For instance, the GNN model achieved an F1 score of 92.1% (appendix 1; Guo et al. (2023), indicating its effectiveness for large-scale to medium-scale generalization (1:10,000 to 1: 50,000). Similarly, Karsznia et al. (2024a) reported high accuracy rates with various ML models, including DTGA achieving 90% accuracy (appendix 1; Karsznia et al. 2024a) for smaller scale map design.

Stroke-based approaches have also shown good results, particularly when applied to medium and small-scale generalization. For example, Weiss and Weibel (2014) demonstrated a mean improvement of 67.88% over basic methods in multiple research areas using enhanced stroke generation techniques (appendix 1; Weiss & Weibel 2014). This approach proved to be effective for the target scale of 1:200 000. In a similar way the AHP implementation to stroke generation proposed by Han et al. (2020) allowed to achieve rather high common stroke ratio of 89% for the same target scale. Stroke generation was also effective both in the research conducted by Li et al. (2020) and Zhang et al. (2017) for road generalization up to 1:50 000 scale. The proposed methods achieved 91.64% and 88.80% maximum similarity respectively.

Mesh-based approaches, such as those employed by Benz and Weibel (2014), are particularly helpful at maintaining structural integrity and satisfying so-called hard constraints. Their extended stroke–mesh combination method enabled to achieve 100% satisfaction concerning hard constraints which included: percentage of roads to be kept, necessary segments that cannot be removed, keeping complete road connectivity, not creating new dead ends, not collapsing the roundabouts, and keeping important POIs accessible (appendix 1; Benz & Weibel 2014). Satisfying these hard constraints makes the approach suitable for medium-scale applications, that is, from 1:10 000 to 1:50 000. Moreover, this method requires minimal manual corrections, which is a significant advantage in practical applications. Similarly, Chen et al. (2009) demonstrated a mean consistency for all test areas at the level of 89.50% with existing maps using mesh density-based selection for scales from 1:10 000 to 1:50 000 (appendix 1; Chen et al. 2009).

Graph-based approaches build on the notion that roads and intersections can be represented as nodes and edges, thereby focusing on topological relationships. For example, Lyu et al. (2022) reported an accuracy of 86% in road-path selection, demonstrating that these methods can perform competitively at scales transitioning from 1:50 000 to 1:100 000 (appendix 1; Lyu et al. 2022). Similarly, Pung et al. (2022) introduced correlation-based measures such as Pearson (ρ = 0.964) and Spearman (R = 0.911) to evaluate how closely the generalized network structure aligns with the source configuration at large scales (appendix 1; Pung et al. 2022). Earlier work by Zhang et al. (2011) highlights the potential of graph-based reasoning employing ego network and weighted ego network analyses. Ego networks consist of a single actor (ego) together with the actors that they are connected to (alters) and all the links among those alters (Evert & Borgatti 2005). Use of such a method allowed to maintain high retention rates (above 90%) of critical road segments in scale-free contexts (Zhang et al. 2011). Compared to older stroke- or mesh-based methods, these graph-centric techniques not only sustain or improve the established accuracy standards, but also enrich the evaluative framework by capturing topological, structural, and relational qualities.

The hybrid approaches that combine stroke and mesh methods have also shown high usefulness. In this scope Touya (2010) proposed an enriched structural selection method that allowed to achieve a road length overlap of 97%, showing the method’s ability to maintain consistency and structural integrity (appendix 1; Touya 2010). This hybrid approach is particularly effective for medium- to small-scale transitions, such as from 1:50 000 to 1:100 000. However, the author pointed out that in urban road networks the road length overlap could be falsely high because many segments in more dense areas have identical lengths. The work of Zhou and Li (2017) supports the application of ML and elements dedicated to graph type solutions (enriching information about roads with centrality measures). In this research high accuracy rates were achieved across multiple target scales, such as 91.55% for 1:200 000 using the MP method (appendix 1; Zhou & Li 2017).