Graphic Simpliﬁcation and Intelligent Adjustment Methods of Road Networks for Navigation with Reduced Precision

: With the rapid development of high-precision road network maps, low-precision road network maps (basic data unrelated to hardware) will need to be directly produced for traditional navigation software from high-precision maps. To do so, large amounts of vector data representing road networks must be simpliﬁed and spatial directional similarity in road networks must be maintained while reducing precision. In this study, an elite strategy genetic algorithm based on the grid model is applied to spatial directional adjustment in road networks for producing road network maps for traditional navigation. Firstly, semantic features and critical vertices are extracted from the road network with high precision. Secondly, some high-precision vertices are eliminated under constraints of the digital navigation map. During this process, the local shape maintenance of the road is considered, and the destruction of the spatial topological relationships is avoided. Thirdly, a genetic algorithm for minimizing the total changes in road azimuths at nodes of road networks is developed to maintain spatial directional relationships while reducing precision. Experimental results and visualization e ﬀ ects on the test data of di ﬀ erent cities show that this method is suitable for generating road network maps for traditional navigation software from high-precision ones.


Introduction
Research on line simplification is valuable in the fields of computer graphics and cartography because over 80% of the elements on maps are lines [1]. The classic Douglas-Peucker (DP) algorithm [2] is essentially a process of recursively extracting valid points; the algorithm considers local as well as global features when selecting points. In the Opheim simplification algorithm [3,4], the maximum and minimum distance constraints determine the search area, thereby eliminating the intermediate original points, except the first and last points, in the search area. Visvalingam and Whyatt [5] firstly proposed an algorithm using the area as a simplified indicator. The area formed by each intermediate original point and its two adjacency efficient points is calculated. Then, the intermediate original point with the smallest effective area can be removed to ensure that the shape of the simplified line changes minimally. However, these basic algorithms do not consider the spatial topological relationship during simplification. On this basis, these algorithms have been improved. Based on the original DP algorithm, Saalfeld [6] detected and removed existing topological conflicts by dynamically updating ( Figure 1). For example, after the precision is directly reduced, the point A' is located at position 3, whereas the azimuth deviation caused by position 2 is smaller. Thus, we focused on two problems: the simplification of roads, the maintenance of the spatial directional relationship between roads, and especially how to minimize the spatial directional change of roads after reducing accuracy. The effective area, vertical distance, buffer radius, etc., are common metrics in independent point algorithms [20], local processing routines [21][22][23][24], or global routines [25][26][27] for line simplification. As a classic line simplification algorithm, the Opheim algorithm can express the distance constraint between vertices in road networks. However, it does not consider other constraints such as turning angles of a road; therefore, for the road simplification, the Opheim algorithm must be improved. distance view and the feature set view, which are the views used for similarity cognition. With the developed theory of complex spatial scenes, other models for the description of the spatial scene and their similarity measures were proposed [32,33]. Yan et al. [34] focused on quantitative expressions of the features of spatial similarity relations in map spaces. Lewis and Egenhofer [35] developed a model for spatial scene description using sequences of detailed spatial intersections and object containment, capturing the connections between complex spatial objects. Buttenfield [36] presented an algorithm that maintains geometric and topological characteristics while progressively transmitting vector coordinates. However, these studies did not consider the precision reduction during road network simplification. Guo et al. [37] proposed an algorithm for the adjustment of the spatial direction relationship at the nodes of road networks only considering the local optimal solution. In this research, we applied an intelligent algorithm to find the global optimal solution in the process of adjusting the spatial direction relationship between roads at the nodes of the road network.  When the data precision is five digits, mathematically, all points on the road must be at the intersection of square grids with one length unit, so the grid model is used in this approach. We used both the improved Opheim algorithm and the spatial directional relationship maintenance of road networks to produce low-precision road network maps that meet the production requirements and When the data precision of roads is lowered from seven to five digits, the spatial directional relationship of simplified roads must be adjusted in accordance with the navigation requirements of road networks because the reduction in data precision causes changes in the spatial directional relationship among roads. The similarity evaluation method is based on the calculation of similarity among directional relationship matrices, as proposed by Goyal and Egenhofer [28]. It is based on the main direction distance calculation of the conceptual neighborhood graph, which is conducive to the evaluation of scene similarity. Considering direction relationships between two objects at different levels of detail, Goyal and Egenhofer [29] extended the model of the direction-relation matrix and studied the models of spatial directional relationship for arbitrary pairs of points in map generalization. Tang et al. [30,31] studied the measurement of graphical similarity among linear roads, and proposed a measuring model for linear spatial data similar based on the integration of the distance view and the feature set view, which are the views used for similarity cognition. With the developed theory of complex spatial scenes, other models for the description of the spatial scene and their similarity measures were proposed [32,33]. Yan et al. [34] focused on quantitative expressions of the features of spatial similarity relations in map spaces. Lewis and Egenhofer [35] developed a model for spatial scene description using sequences of detailed spatial intersections and object containment, capturing the connections between complex spatial objects. Buttenfield [36] presented an algorithm that maintains geometric and topological characteristics while progressively transmitting vector coordinates. However, these studies did not consider the precision reduction during road network simplification. Guo et al. [37] proposed an algorithm for the adjustment of the spatial direction relationship at the nodes of road networks only considering the local optimal solution. In this research, we applied an intelligent algorithm to find the global optimal solution in the process of adjusting the spatial direction relationship between roads at the nodes of the road network.
When the data precision is five digits, mathematically, all points on the road must be at the intersection of square grids with one length unit, so the grid model is used in this approach. We used both the improved Opheim algorithm and the spatial directional relationship maintenance of road networks to produce low-precision road network maps that meet the production requirements and visual effects. We especially focused on the change in road shape and the spatial directional relationship at nodes. The following contributions are provided by this work: (1) An improved Opheim algorithm combined with the local shape maintenance of roads.
(2) In the process of both road simplification and adjustment of the spatial directional relationship, the consistency maintenance of topological relationships is considered to avoid topological errors.
(3) If a node is directly connected to any point that is not one of the other nodes, this node is called a general node (Figure 2). When road data precision is reduced, the spatial directional relationship among road segments connected at every general node is adjusted optimally; the change in the spatial directional relationship at every general node is the smallest after reducing precision. relationship, the consistency maintenance of topological relationships is considered to avoid topological errors.
(3) If a node is directly connected to any point that is not one of the other nodes, this node is called a general node (Figure 2). When road data precision is reduced, the spatial directional relationship among road segments connected at every general node is adjusted optimally; the change in the spatial directional relationship at every general node is the smallest after reducing precision.
(4) If two or more nodes are directly connected, these nodes are called special nodes ( Figure 2). The genetic algorithm is applied with the elite strategy (e-GA) to the adjustment of the spatial directional relationship at special nodes, which helps to intelligently find the global optimal solution. The total variation in the spatial directions of road segments connected at a subgroup of special nodes can be as small as possible after the precision of road data in road network generalization is reduced.
The main steps of our method involve two workflows for road data handling ( Figure 2). The first one is road simplification while maintaining the spatial direction relationship of the roads connected to the vertices, which includes the constraints in Section 2, the method for extracting road topographical features in Section 3 and the improved Opheim algorithm in Section 4. The second workflow is the maintenance of the spatial directional relationship of a subgroup of the road network connected with special nodes or a general node, which is described in Section 5. We analyze the results of the experiments in Section 6, and provide our conclusions in Section 7.

Data Model of Roads
In the road data model for navigation, a set of points {P0, …, Pi, …, Pm − 1} represents a series of sequential points on the road. Nodes Vj = P0 and Vk = Pm − 1 represent the endpoint or the 3D virtual intersection point of the road, and point Pi (i ≠ 0, i ≠ m − 1) represents the vertex of the road, as shown in Figure 3a. 3D virtual intersection points and endpoints are distinguished in accordance with the attribute flag of points. If flag = 0, then it is the 3D virtual intersection point at which two projected roads intersect in 2D space but not in 3D space. If flag = 2, then it is the endpoint; if flag = 1, this indicates the vertex.
The road network in 2D space can be viewed as a graph structure G composed of a set of nodes V with an attribute flag and a set of edges E and is denoted as G = (V, E), as shown in Figure 3b. In the road network in this study, let an edge be the road section between two nodes and road segment be (4) If two or more nodes are directly connected, these nodes are called special nodes ( Figure 2). The genetic algorithm is applied with the elite strategy (e-GA) to the adjustment of the spatial directional relationship at special nodes, which helps to intelligently find the global optimal solution. The total variation in the spatial directions of road segments connected at a subgroup of special nodes can be as small as possible after the precision of road data in road network generalization is reduced.
The main steps of our method involve two workflows for road data handling ( Figure 2). The first one is road simplification while maintaining the spatial direction relationship of the roads connected to the vertices, which includes the constraints in Section 2, the method for extracting road topographical features in Section 3 and the improved Opheim algorithm in Section 4. The second workflow is the maintenance of the spatial directional relationship of a subgroup of the road network connected with special nodes or a general node, which is described in Section 5. We analyze the results of the experiments in Section 6, and provide our conclusions in Section 7.

Data Model of Roads
In the road data model for navigation, a set of points {P 0 , . . . , P i , . . . , P m − 1 } represents a series of sequential points on the road. Nodes V j = P 0 and V k = P m − 1 represent the endpoint or the 3D virtual intersection point of the road, and point P i (i 0, i m − 1) represents the vertex of the road, as shown in Figure 3a. 3D virtual intersection points and endpoints are distinguished in accordance with the attribute flag of points. If flag = 0, then it is the 3D virtual intersection point at which two projected roads intersect in 2D space but not in 3D space. If flag = 2, then it is the endpoint; if flag = 1, this indicates the vertex. the section between two sequential points. In Figure 3a, road segment P1P2 is associated with node Vj, and points P2 and P3 are adjacent points. Therefore, three types of points exist in the road network: the intersection points, endpoints and vertices. Edges are connected at endpoints or intersection points. Thus, nodes may be endpoints or intersection points, as shown in Figure 3b. Let Vi∈V (G), and the number of edges associated with point Vi in G is called the degree of Vi [38]. In Figure 3b, the degree of node V1 is 3, and the degree of node V2 is 4.

Constraints on Road Generalization
The planar coordinates of points in roads can be represented by (ψ, r), where ψ is the longitude, and r is the latitude. For example, (15.1234567°, 16.1234567°) represents the location of a planar point, where 15.1234567° is the longitude and 16.1234567° is the latitude. The value 15.1234567° is a number with seven decimal places. When reducing road data precision, (15.1234567°, 16.1234567°) must be rounded up to five decimal places, e.g., (15.12346°, 16.12346°). The grid model is used to represent the different precision roads. In Figure 1, the blue line represents a road with seven-digit precision and the orange line represents the same road with five-digit precision. When a road network with seven-digit precision is generalized to five-digit precision, the four constraints on road generalization are as follows: (1) Position precision constraint: The intersection points and endpoints cannot be eliminated, whereas the vertices may be eliminated. During local shape maintenance, the constraint of point movement is d ≤ √2δ, where d is the distance from the point with seven-digit coordinates to that with five-digit coordinates and δ is the coordinate precision defined by users.
(2) Distance constraint: The threshold of the minimum distance between two adjacent points on a road is set to ε, and ε ≤ |Pi − Pi + 1| ≤ k × ε (ε > √2, k > 1, which is defined by users, and Pi and Pi + 1 are two adjacent points).
(3) Constraints for spatial directional similarity: The azimuth of the road segment is the angle between it and the northward direction. In Figure 4a, Ai is the azimuth of road segment ViPi + 1 with seven-digit precision and A'i is the azimuth of road segment ViP'i + 1 with five-digit precision. The spatial directional relationship between two sequential road segments on a road is described by the turning angle. In Figure 4b, Ti is the turning angle of two road segments at vertex Pi with seven-digit precision, and T'i is the turning angle of the same road segment at vertex Pi with five-digit precision.

A´i
Ai Vi Pi+1 P´i+1 The road network in 2D space can be viewed as a graph structure G composed of a set of nodes V with an attribute flag and a set of edges E and is denoted as G = (V, E), as shown in Figure 3b. In the road network in this study, let an edge be the road section between two nodes and road segment be the section between two sequential points. In Figure 3a, road segment P 1 P 2 is associated with node V j , and points P 2 and P 3 are adjacent points. Therefore, three types of points exist in the road network: the intersection points, endpoints and vertices. Edges are connected at endpoints or intersection points. Thus, nodes may be endpoints or intersection points, as shown in Figure 3b. Let V i ∈V (G), and the number of edges associated with point V i in G is called the degree of V i [38]. In Figure 3b, the degree of node V 1 is 3, and the degree of node V 2 is 4.

Constraints on Road Generalization
The planar coordinates of points in roads can be represented by (ψ, r), where ψ is the longitude, and r is the latitude. For example, (15. Figure 1, the blue line represents a road with seven-digit precision and the orange line represents the same road with five-digit precision. When a road network with seven-digit precision is generalized to five-digit precision, the four constraints on road generalization are as follows: (1) Position precision constraint: The intersection points and endpoints cannot be eliminated, whereas the vertices may be eliminated. During local shape maintenance, the constraint of point movement is d ≤ √ 2δ, where d is the distance from the point with seven-digit coordinates to that with five-digit coordinates and δ is the coordinate precision defined by users.
(2) Distance constraint: The threshold of the minimum distance between two adjacent points on a road is set to ε, and ε ≤ |P i − P i + 1 | ≤ k × ε (ε > √ 2, k > 1, which is defined by users, and P i and P i + 1 are two adjacent points).
(3) Constraints for spatial directional similarity: The azimuth of the road segment is the angle between it and the northward direction. In Figure 4a, A i is the azimuth of road segment V i P i+1 with seven-digit precision and A' i is the azimuth of road segment V i P i+1 with five-digit precision. The spatial directional relationship between two sequential road segments on a road is described by the turning angle. In Figure 4b, T i is the turning angle of two road segments at vertex P i with seven-digit precision, and T' i is the turning angle of the same road segment at vertex P i with five-digit precision.
The general node V m in the road network is unaffected by the adjustment of other nodes in the maintenance of the spatial directional relationship. In other words, no other node is connected with it (node V m , Figure 5a). As shown in Figure 5a, all possible five-digit positions of all road segments associated with node V m are first calculated. We can then choose the minimum sum in all possible azimuth variations. The smaller the total azimuth variation, the more similar the spatial direction. The similarity of every azimuth variation (τ) at node V m can be calculated using Equation (1), where η 1 is the threshold of the azimuth variation of a single road segment between two adjacent points: where SimilarityV m is the azimuth similarity of all road segments associated with node V m , M is the degree of node V m , A i is the azimuth of road segment l i (l i is associated with node V m ) with seven-digit precision, A' i is the azimuth of l i with five-digit precision, and j and k represent the optional positions of the node and its adjacency point under five-digit precision on this road segment l i , respectively, which range from integer 0 to 3.
represent the different precision roads. In Figure 1, the blue line represents a road with seven-digit precision and the orange line represents the same road with five-digit precision. When a road network with seven-digit precision is generalized to five-digit precision, the four constraints on road generalization are as follows: (1) Position precision constraint: The intersection points and endpoints cannot be eliminated, whereas the vertices may be eliminated. During local shape maintenance, the constraint of point movement is d ≤ √2δ, where d is the distance from the point with seven-digit coordinates to that with five-digit coordinates and δ is the coordinate precision defined by users.
(2) Distance constraint: The threshold of the minimum distance between two adjacent points on a road is set to ε, and ε ≤ |Pi − Pi + 1| ≤ k × ε (ε > √2, k > 1, which is defined by users, and Pi and Pi + 1 are two adjacent points).
(3) Constraints for spatial directional similarity: The azimuth of the road segment is the angle between it and the northward direction. In Figure 4a, Ai is the azimuth of road segment ViPi + 1 with seven-digit precision and A'i is the azimuth of road segment ViP'i + 1 with five-digit precision. The spatial directional relationship between two sequential road segments on a road is described by the turning angle. In Figure 4b, Ti is the turning angle of two road segments at vertex Pi with seven-digit precision, and T'i is the turning angle of the same road segment at vertex Pi with five-digit precision.

A´i
Ai Vi Pi+1 P´i+1  (4) Constraints for topological consistency: If vertex Pi is eliminated or moved, then the original topological relationship may be destroyed, as shown in Figure 7. In the process of vertex elimination, we can judge whether the neighborhood points are in the triangle being removed. We then know whether the elimination of point Pi destroys the topology, as shown in Figure 7a, i.e., self-intersecting polyline L1 or polyline L1 intersecting with polyline L2 when removing point Pi. In the maintenance of the spatial directional relationship, as shown in Figure 7b, node Vi and any two If a node is connected with at least another node by a straight segment, these nodes are called special nodes (for example, Node 1~4, 6, 8~10 in Figure 6). Special nodes connected with adjacent ones must be regarded as an entirety called local subgroup G. For local subgroups G in road networks, the spatial directional similarity of every possible changed situation (τ) of the road segments connected at nodes in G can be calculated using Equation (2). The five-digit locations of the special nodes are selected when SimilarityG(τ) is the largest in all possible states.
where SimilarityG denotes the azimuth similarity of road segments in set L associated with special nodes in G after reducing precision. Set L stores the road segments connected to each node in G, where ISPRS Int. J. Geo-Inf. 2020, 9, 490 7 of 23 the data is non-duplicated. N is the number of road segments in L; B i is the azimuth of l i (l i ∈L) with seven-digit precision; B' i is the azimuth of l i with five-digit precision; and j and k represent the optional positions of the two nodes associated with l i , ranging from integer 0 to 3.
five-digit node Maintenance of road spatial directional relationship. Road segments associated with (a) a eneral node and (b) the vertex.

4) Constraints for topological consistency: If vertex
Pi is eliminated or moved, then the ori ogical relationship may be destroyed, as shown in Figure 7. In the process of v ation, we can judge whether the neighborhood points are in the triangle being removed know whether the elimination of point Pi destroys the topology, as shown in Figure 7a tersecting polyline L1 or polyline L1 intersecting with polyline L2 when removing point P aintenance of the spatial directional relationship, as shown in Figure 7b, node Vi and any iated road segments (e.g., l1 and l2) form a triangular sector S. We judge whether tment leads to the loss of topological consistency by judging whether the points (e.g., poin tor S are in sector S' formed by node V'i and two road segments with five-digit precision aintenance of the spatial directional relationship. Simultaneously, the relative positions of l4 remain unchanged.  Let P i be the vertex on the road, and the spatial directional relationship at vertex P i is measured by the turning angle. As shown in Figure 5b, P i − 1 , P i and P i + 1 can be located at four optional positions when reducing precision. When the turning angle difference from seven-to five-digit precision at vertex P i is the smallest, the similarity of spatial directional relationship at vertex P i is the largest, as shown in Equation (3). η 2 is the threshold of turning angle variation at vertex P i .
where SimilarityP i indicates the spatial direction similarity of the road at vertex P i , i (i∈{0,1, . . . , m − 1}) represents the ith vertex on a road, T i is the turning angle at the ith vertex before the local shape is adjusted, T' i is the turning angle at the ith vertex with five-digit precision after the local shape is adjusted, m is the number of vertices included on the road, and j represents one of the optional positions of vertex P' i and is an integer from 0 to 3. (4) Constraints for topological consistency: If vertex P i is eliminated or moved, then the original topological relationship may be destroyed, as shown in Figure 7. In the process of vertex elimination, we can judge whether the neighborhood points are in the triangle being removed. We then know whether the elimination of point P i destroys the topology, as shown in Figure 7a, i.e., self-intersecting polyline L 1 or polyline L 1 intersecting with polyline L 2 when removing point P i . In the maintenance of the spatial directional relationship, as shown in Figure 7b, node V i and any two associated road segments (e.g., l 1 and l 2 ) form a triangular sector S. We judge whether this adjustment leads to the loss of topological consistency by judging whether the points (e.g., point P j ) in sector S are in sector S' formed by node V' i and two road segments with five-digit precision after the maintenance of the spatial directional relationship. Simultaneously, the relative positions of l 1 , l 2 , l 3 , and l 4 remain unchanged.
self-intersecting polyline L1 or polyline L1 intersecting with polyline L2 when removing point Pi. In the maintenance of the spatial directional relationship, as shown in Figure 7b, node Vi and any two associated road segments (e.g., l1 and l2) form a triangular sector S. We judge whether this adjustment leads to the loss of topological consistency by judging whether the points (e.g., point Pj) in sector S are in sector S' formed by node V'i and two road segments with five-digit precision after the maintenance of the spatial directional relationship. Simultaneously, the relative positions of l1, l2,

Extraction of 3D Terrain Feature
Slope describes the topographical feature, which is important information in navigation for a 3D road. The positive and negative signs of the slope reflect uphill and downhill, respectively. Thresholds α 1 and α 2 are given (α 1 < α 2 ), where α 1 represents the upper limit of the slope of the flat road, and α 2 represents the lower limit of the steep slope. Slope i,1 is the slope of road segment P i−1 P i , and Slope i,2 is the slope of road segment P i P i+1 . If and only if |Slope i,1 | ≤ α 1 and |Slope i,2 | ≤ α 1 , then a partial flat road at P i exists; otherwise, slope variation occurs.
The slope variation at vertex P i can be expressed by ∆Slope i , where ∆Slope i = |Slope i,1 − Slope i,2 |, which is related to the two road segments associated with P i . Given threshold β when ∆Slope i > β, the slope variation at P i is large and P i is defined as a critical terrain point. The six types of critical terrain points (types 1 -6 ) on a road are shown in Figure 8. Axis y represents the elevation (z) of vertices on a road and axis x represents the length (x) of the road section from a vertex to the node along the road. Origin O is a node of the road. The corresponding identification methods of the critical terrain points proposed in this study are shown in Equation (4).
Slope describes the topographical feature, which is important information in navigation for a road. The positive and negative signs of the slope reflect uphill and downhill, respectively. esholds α1 and α2 are given (α1 < α2), where α1 represents the upper limit of the slope of the flat , and α2 represents the lower limit of the steep slope. Slopei,1 is the slope of road segment Pi-1Pi , Slopei,2 is the slope of road segment PiPi + 1 . If and only if |Slopei,1| ≤ α1 and |Slopei,2| ≤ α1, then a ial flat road at Pi exists; otherwise, slope variation occurs.
The slope variation at vertex Pi can be expressed by ΔSlopei, where ΔSlopei = |Slopei,1 − Slopei,2|, ch is related to the two road segments associated with Pi. Given threshold β when ΔSlopei > β, the e variation at Pi is large and Pi is defined as a critical terrain point. The six types of critical terrain ts (types ①-⑥) on a road are shown in Figure 8. Axis y represents the elevation (z) of vertices on ad and axis x represents the length (x) of the road section from a vertex to the node along the . Origin O is a node of the road. The corresponding identification methods of the critical terrain ts proposed in this study are shown in Equation (4).

Extraction of a Long, Straight and Flat Road
For road network maps, long, straight, and flat road sections provide important semantic information. The length threshold of a long, straight, and flat line is defined as γ, and a strip with a small width µ [23] is used to extract it from the polyline, as shown in Figure 9. Initial point P j is at the center of the strip. The initial direction of the strip is parallel to the road segment composed of ISPRS Int. J. Geo-Inf. 2020, 9, 490 9 of 23 the initial point and its next point. The strip shifts over along the initial direction until the strip is just away from the polyline. Point set {P j , P j + 1 , . . . , P j + m }, which the strip has passed, forms a straight line. If the length L of the straight line is larger than γ, then the two endpoints of this road section from P j to P j + m are marked as critical topographical points for a long, straight, and flat road section. P j + m is the next initial point of the remaining polyline on the road, and the same process is repeated until the strip passes the end of the polyline. Finally, all long, straight, and flat road sections are extracted.
For road network maps, long, straight, and flat road sections provide important semantic information. The length threshold of a long, straight, and flat line is defined as γ, and a strip with a small width μ [23] is used to extract it from the polyline, as shown in Figure 9. Initial point Pj is at the center of the strip. The initial direction of the strip is parallel to the road segment composed of the initial point and its next point. The strip shifts over along the initial direction until the strip is just away from the polyline. Point set {Pj, Pj + 1, …, Pj + m}, which the strip has passed, forms a straight line. If the length L of the straight line is larger than γ, then the two endpoints of this road section from Pj to Pj + m are marked as critical topographical points for a long, straight, and flat road section. Pj + m is the next initial point of the remaining polyline on the road, and the same process is repeated until the strip passes the end of the polyline. Finally, all long, straight, and flat road sections are extracted.

Opheim Algorithm
As shown in Figure 10, the Opheim algorithm defines the search region with the greater width µ as the strip in Figure 9, as well as a minimum and maximum distance constraint. All vertices in the search area whose radius is the minimum distance are eliminated. Except for the last vertex (P 3 in Figure 10 left) within the search area whose radius is the maximum distance, other vertices in this search area are also eliminated. Then, this undeleted vertex (P 3 in Figure 10, right) is the center of the next search area, and the vertex P 5 in Figure 10 right can be found. The search area is moved along the original line until another endpoint of this line is selected.

Opheim Algorithm
As shown in Figure 10, the Opheim algorithm defines the search region with the greater width μ as the strip in Figure 9, as well as a minimum and maximum distance constraint. All vertices in the search area whose radius is the minimum distance are eliminated. Except for the last vertex (P3 in Figure 10 left) within the search area whose radius is the maximum distance, other vertices in this search area are also eliminated. Then, this undeleted vertex (P3 in Figure 10, right) is the center of the next search area, and the vertex P5 in Figure 10 right can be found. The search area is moved along the original line until another endpoint of this line is selected.

Improved Opheim Algorithm
The elimination of vertices on a long, straight, and flat road is simple and should not destroy the topological relationship. When a long, straight, and flat road section is searched, the two endpoints of this road section should not be eliminated. The vertices between the two endpoints are usually eliminated, but some vertices between these endpoints must be retained in accordance with navigational requirements; the distance between two retained sequential vertices should be larger than a threshold defined by the road network map model. In this study, the Opheim line simplification algorithm [3] was used to eliminate the vertices between two endpoints of a long, straight, and flat road section.
The elimination of vertices on curved roads is similar to that on straight roads; however, it is more complicated because all critical points and various possible situations must be considered. The minimum distance constraint threshold is ε; the search area is determined with the initial point as the center of the circle and the distance constraint threshold ε as the radius. Before vertex elimination, important critical vertices and nodes in the road network for navigation should not be eliminated. The attribute value Ranki of every point in road topographical feature points is defined as Ranki = 1. Other vertices' Ranki = 0. In the process of vertex elimination, if the elimination of vertex Pi causes topological conflicts, then Ranki = 2. The judgment method of topological conflicts was described in Section 2. We

Improved Opheim Algorithm
The elimination of vertices on a long, straight, and flat road is simple and should not destroy the topological relationship. When a long, straight, and flat road section is searched, the two endpoints of this road section should not be eliminated. The vertices between the two endpoints are usually eliminated, but some vertices between these endpoints must be retained in accordance with navigational requirements; the distance between two retained sequential vertices should be larger than a threshold defined by the road network map model. In this study, the Opheim line simplification algorithm [3] was used to eliminate the vertices between two endpoints of a long, straight, and flat road section.
The elimination of vertices on curved roads is similar to that on straight roads; however, it is more complicated because all critical points and various possible situations must be considered. The minimum distance constraint threshold is ε; the search area is determined with the initial point as the center of the circle and the distance constraint threshold ε as the radius. Before vertex elimination, important critical vertices and nodes in the road network for navigation should not be eliminated. The attribute value Rank i of every point in road topographical feature points is defined as Rank i = 1. Other vertices' Rank i = 0. In the process of vertex elimination, if the elimination of vertex P i causes topological conflicts, then Rank i = 2. The judgment method of topological conflicts was described in Section 2.
We suppose a road R = {P 0 , P 1 , . . . , P i , . . . , P m − 1 }, where m represents the number of vertices. Let P i−1 be the initial point. The length of P i−1 P i is represented as D i,1 , and the length of P i P i+1 is represented as D i,2 . Each vertex P i on the road has an attribute Remove i , which is used to indicate whether the vertex will be eliminated. Remove i = 0 means vertex P i will neither be eliminated nor marked as a problem point. Remove i = 1 means vertex P i will be eliminated. Remove i = 2 means vertex P i will not be eliminated but marked as a problem point to be handled further because the elimination of vertex P i causes a topological conflict. Given the angle threshold θ, if the absolute value of turning angle T i at vertex P i is smaller than the threshold θ, then vertex P i should be eliminated. The algorithm of determining whether vertex P i should be eliminated is illustrated in Figure 11, and the pseudocode is given as Algorithm 1. if Rank i = 2 5.

Local Shape Maintenance While Reducing Precision
As the precision of road network data is lowered, the road shape should change, as shown in Figure 1. In the grid model, each point with seven-digit precision is converted into one with five-digit precision, which has four optional positions. The change in the turning angle at point Pi is used in this study to describe the change in road shape. The spatial direction similarity SimilarityPi at point Pi represents the local shape similarity. The angle threshold θ was given, and the absolute value of turning angle Ti at vertex Pi within the threshold θ represents a straight line. The road directions at vertex Pi have two other possibilities: if Ti > θ, then turn left; if Ti < −θ, then turn right.
Local shape maintenance should meet the following two conditions: (1) the change in direction is small and (2) the position is optimal. An optimal position means that selecting the position of its four optional positions of vertex Pi can make the local shape the most similar to the original shape, Figure 11. Vertex elimination on a curved road. (a) D i , 1

Local Shape Maintenance While Reducing Precision
As the precision of road network data is lowered, the road shape should change, as shown in Figure 1. In the grid model, each point with seven-digit precision is converted into one with five-digit precision, which has four optional positions. The change in the turning angle at point P i is used in this study to describe the change in road shape. The spatial direction similarity SimilarityP i at point P i represents the local shape similarity. The angle threshold θ was given, and the absolute value of turning angle T i at vertex P i within the threshold θ represents a straight line. The road directions at vertex P i have two other possibilities: if T i > θ, then turn left; if T i < −θ, then turn right.
Local shape maintenance should meet the following two conditions: (1) the change in direction is small and (2) the position is optimal. An optimal position means that selecting the position of its four optional positions of vertex P i can make the local shape the most similar to the original shape, which maximizes the value of SimilarityP i . As shown in Figure 4b, the optimal location of P i − 1 is supposed to have been selected. The optimal location of vertex P i is related not only to vertex P i − 1 but also to vertex P i + 1 . The selection process of its optimal location is as follows: (1) The impossible position of vertex P i is removed in accordance with SimilarityP i − 1 under five-digit precision. Firstly, the positions with five-digit precision can satisfy the precision constraint and cannot cause the topological error. Secondly, SimilarityP i − 1 can be used to exclude particularly poor locations based on navigational constraints.
(2) The optimal position of vertex P i is selected on the basis of SimilarityP i . After precision is reduced, 16 possible connection situations theoretically exist for vertices P i and P i + 1 . After eliminating impossible positions with five-digit precision, the position with the smallest SimilarityP i in the possible positions with five-digit precision of vertex P i is selected as optimal.
Theoretically, during the local shape maintenance of the road, when topological errors occur in all optional positions of a vertex, the position that makes the adjusted road closest to the original road shape will be selected, and topological errors will be marked in the result. After road shape maintenance, road vertices that do not satisfy the constraints under the five-digit precision should be eliminated. The detailed method was described in Section 4.1.

Adjustment of Spatial Directional Relationship at a General Node
If a node is connected with a straight road, then this node is not a general node and is referred to as a special node here. This straight road only consists of two nodes. Thus, the general node is only adjacent to vertices. The situations of spatial directional relationship maintenance at a general node are shown in Figure 12. The function of representing the best situation in the process of spatial relationship maintenance at any general node is: max SimilarityV m (τ). (5) also to vertex Pi + 1. The selection process of its optimal location is as follows: (1) The impossible position of vertex Pi is removed in accordance with SimilarityPi − 1 under five-digit precision. Firstly, the positions with five-digit precision can satisfy the precision constraint and cannot cause the topological error. Secondly, SimilarityPi − 1 can be used to exclude particularly poor locations based on navigational constraints.
(2) The optimal position of vertex Pi is selected on the basis of SimilarityPi. After precision is reduced, 16 possible connection situations theoretically exist for vertices Pi and Pi + 1. After eliminating impossible positions with five-digit precision, the position with the smallest SimilarityPi in the possible positions with five-digit precision of vertex Pi is selected as optimal.
Theoretically, during the local shape maintenance of the road, when topological errors occur in all optional positions of a vertex, the position that makes the adjusted road closest to the original road shape will be selected, and topological errors will be marked in the result. After road shape maintenance, road vertices that do not satisfy the constraints under the five-digit precision should be eliminated. The detailed method was described in Section 4.1.

Adjustment of Spatial Directional Relationship at a General Node
If a node is connected with a straight road, then this node is not a general node and is referred to as a special node here. This straight road only consists of two nodes. Thus, the general node is only adjacent to vertices. The situations of spatial directional relationship maintenance at a general node are shown in Figure 12. The function of representing the best situation in the process of spatial relationship maintenance at any general node is: The algorithm is described as follows: (1) When the precision is lowered, all possible connection situations between node Vm and its adjacent vertices are considered, and the azimuth deviations of each connection situation (τ) are calculated.
(2) The connection situations that are inconsistent with the original topological relationship and the connection situations in which the azimuth changes are greater than the given threshold η1 are excluded, as shown in Figure 12a.
(3) The remaining connection situations are compared with one another in accordance with SimilarityVm(τ). As depicted in Figure 12b, we choose the optimal situation under five-digit precision.
In this process, if topological errors occur in all adjustment schemes, the optimal scheme selected from them will also cause topological errors, and topological errors will be marked in the result. The algorithm is described as follows: (1) When the precision is lowered, all possible connection situations between node V m and its adjacent vertices are considered, and the azimuth deviations of each connection situation (τ) are calculated.
(2) The connection situations that are inconsistent with the original topological relationship and the connection situations in which the azimuth changes are greater than the given threshold η 1 are excluded, as shown in Figure 12a.
(3) The remaining connection situations are compared with one another in accordance with SimilarityV m (τ). As depicted in Figure 12b, we choose the optimal situation under five-digit precision.
In this process, if topological errors occur in all adjustment schemes, the optimal scheme selected from them will also cause topological errors, and topological errors will be marked in the result.

Difference between Two Types of Adjustments
In Figure 13a, nodes V 1 , V 2 , V 3 , and V 4 are special nodes. On the basis of Equation (5), if the spatial directional relationship at node V 1 is adjusted first, the result of the partial region is as shown in Figure 13b. If the spatial directional relationship at V 2 is adjusted first, the result of the partial region is as shown in Figure 13c. The special nodes directly affect one another. The results of spatial directional maintenance are not only related to the threshold but also close to the processing order of special nodes. For the local subgroups of road networks directly formed by special nodes and their adjacency vertices, the results of spatial directional adjustment may not minimize the overall change in the spatial directional relationship. To solve this problem, we focused on road segments connected with special nodes in the local subgroups of road networks and constructed an intelligent algorithm to achieve the optimal combination while reducing the precision of the road network.

Difference between Two Types of Adjustments
In Figure 13a, nodes V1, V2, V3, and V4 are special nodes. On the basis of Equation (5), if the spatial directional relationship at node V1 is adjusted first, the result of the partial region is as shown in Figure 13b. If the spatial directional relationship at V2 is adjusted first, the result of the partial region is as shown in Figure 13c. The special nodes directly affect one another. The results of spatial directional maintenance are not only related to the threshold but also close to the processing order of special nodes. For the local subgroups of road networks directly formed by special nodes and their adjacency vertices, the results of spatial directional adjustment may not minimize the overall change in the spatial directional relationship. To solve this problem, we focused on road segments connected with special nodes in the local subgroups of road networks and constructed an intelligent algorithm to achieve the optimal combination while reducing the precision of the road network.

Elite Strategy Genetic Algorithm for Spatial Direction Adjustment
The special nodes directly connected in a road network must be organized as a community structure for spatial directional relationship maintenance called a local subgroup. Many methods for extracting community structure [39] or generalizing road networks [40] have been studied, such as a spectral algorithm for community detection based on the characteristic matrix of the network [41] and detecting communities by compressing the description of information flows on networks [42]. We used cyclic iteration in this work, and the special nodes on directly connected road networks are continuously integrated into the same local subgroup G. The adjacency vertices of road segments associated with nodes are also added to the local subgroup G as 'virtual nodes' to adjust them.

Elite Strategy Genetic Algorithm for Spatial Direction Adjustment
The special nodes directly connected in a road network must be organized as a community structure for spatial directional relationship maintenance called a local subgroup. Many methods for extracting community structure [39] or generalizing road networks [40] have been studied, such as a spectral algorithm for community detection based on the characteristic matrix of the network [41] and detecting communities by compressing the description of information flows on networks [42]. We used cyclic iteration in this work, and the special nodes on directly connected road networks are continuously integrated into the same local subgroup G. The adjacency vertices of road segments associated with nodes are also added to the local subgroup G as 'virtual nodes' to adjust them.
When the precision of road data is lowered from seven to five digits, each node or its adjacency vertex has four alternatives. 4 N optional situations exist for a subgroup of special nodes, where N is the number of nodes and their adjacency points in G. When N is relatively large, the optimal solution is impossible to find through the exhaustive method. A genetic algorithm with the elite strategy [43] was used in this study to solve the problem. Elite strategy means if the fitness of an individual in the past populations is larger than that of every individual in the current population, this string is preserved in the current generation. On this basis, we add a fixed-capacity elite retention library to retain the best individuals up to the current generation. The problem is abstracted to determine the arrangement and combination of certain items satisfying the constraints. The elite strategy can ensure that the optimal individual is not destroyed. Intelligent algorithms are widely used in map generalization [44][45][46][47]. The process for the spatial directional relationship adjustment of local subgroups of a road network based on the genetic algorithm with elite strategy (e-GA) is described as: (1) Gene coding: Each node or its adjacency vertex in the local subgroups of road network G is abstracted as a gene, and all the genes contained constitute a chromosome or an individual, as shown in Figure 6. In the inheritance process, all individuals of each generation form a population. From sevento five-digit precision, each node or its adjacency vertex has four optional locations. For convenience, real coding is used to directly encode gene phenotypes. The number of nodes in the local subgroups of a road network and their adjacency vertices is assumed as N, and the solution space can be modeled as a chromosome containing N genes as where S represents the solution space, i is the number of genes, and C i is a series of alleles located on the ith gene.
(2) Adaptability evaluation. The fitness function is the basis for evaluating the quality of the results. This function is important for maintaining the spatial directional relationship. The calculation method was shown in Equation (2). The greater the fitness, the better the maintenance effect of the spatial directional relationship. The optimal solution must satisfy the maximum fitness and constraint. The constraint is (3) Genetic manipulation refers to selection, crossover, and variation. Holland's roulette wheel selection [48] is the classic genetic selection operator. In this process, the better the individuals selected for mating, the greater the probability excellent genes will be inherited. For a population with number N, Group = {S 0 , S 2 , . . . , S N − 1 }. The fitness value of an individual is f (S i ). The probability P i selected is A crossover operation is used to exchange some genes between two individuals with a certain probability to ensure the robustness of the population. The roulette algorithm was used in this study to randomly select two individuals. According to the gene size, let k 1 and k 2 be two random cross points, and k 1 and k 2 range from 0 to N − 1. Chromosomes are divided into three loci, and a sequence of gene fragments is exchanged on the basis of the crossover probability P c (0 < P c < 1). The crossing process of two individuals for a subgroup of special nodes is depicted in Figure 14. (4) Elite strategy. Individuals with a high fitness value in the current population are kept in the elite library with number K. If their fitness value is no better than in the next generation, then individuals retained in the elite library are directly copied to the next generation and the corresponding number of the worst individuals in the next generation is replaced. The individuals in the next generation are sorted in descending order in accordance with the fitness value, and the former K individuals with large fitness values are used to update the elite retention library. When the iteration stops, the best retained individuals are approximated to the global optimal solution. Under the elite retention strategy, the genetic algorithm can converge to the global optimal solution of the problem with probability 1.
(5) Optimal solution judgment. In this study, the e-GA determines whether the cycle terminates by judging the convergence degree of the population, which is based on the change in the fitness value of the individuals retained in the elite retention library in each generation. The maximum generation is set to avoid searching for an approximate solution after the pre-accepted solution is obtained. The criteria for whether the population converges are as follows: a. Among K best individuals in the elite retention library, the one with the smallest fitness value is better than the best individual among those in the next t generation.
b. The maximum generation is reached. For the final K best individuals, the individuals with the smallest distance change are selected as the optimal solution from the individuals with minimal difference in fitness values. The calculation of distance change ΔD(j) is shown in Equation (9). The nodes and their adjacency vertices in the local subgroups of road network G are adjusted on the basis of the optimal scheme. In this process, the optimal scheme may cause topological errors. However, these topological errors will be marked in the result: where i is the ith node or the adjacency vertex in G, j is its position, and ΔD is the sum of distance changes of the nodes and their adjacency vertices in G. Di is the position of the ith node or the adjacency vertex in G under seven-digit precision and D'i is the position of the ith node or the adjacency vertex in G under five-digit precision.

Experiments Design and Study Area
The proposed algorithm was implemented in C#, and the software developed for it can produce a road network map with five-digit precision from one with seven-digit precision. The seven-digit road network data (Advanced Driving Assistance System (ADAS) in specific formats) of a region in Chongqing, China, which was provided by NavInfo Co., Ltd. (Beijing, China), which included 53,525 roads, 106,537 endpoints, 8619 intersection points, and 118,327 vertices, were selected as experimental data. The study area is shown in Figure 15. The mutation operation guarantees the diversity of the population and belongs to a small-probability event. Multiple single-point mutations were used in this study to select individuals from roulette selection and crossover for random variation. The number of mutations is n (0 ≤ n ≤ N), and each time, the mutation is based on mutation probability P m . Multiple mutations can maintain individual differences in the population and prevent premature convergence of the program.
(4) Elite strategy. Individuals with a high fitness value in the current population are kept in the elite library with number K. If their fitness value is no better than in the next generation, then individuals retained in the elite library are directly copied to the next generation and the corresponding number of the worst individuals in the next generation is replaced. The individuals in the next generation are sorted in descending order in accordance with the fitness value, and the former K individuals with large fitness values are used to update the elite retention library. When the iteration stops, the best retained individuals are approximated to the global optimal solution. Under the elite retention strategy, the genetic algorithm can converge to the global optimal solution of the problem with probability 1.
(5) Optimal solution judgment. In this study, the e-GA determines whether the cycle terminates by judging the convergence degree of the population, which is based on the change in the fitness value of the individuals retained in the elite retention library in each generation. The maximum generation is set to avoid searching for an approximate solution after the pre-accepted solution is obtained. The criteria for whether the population converges are as follows: a. Among K best individuals in the elite retention library, the one with the smallest fitness value is better than the best individual among those in the next t generation.
b. The maximum generation is reached. For the final K best individuals, the individuals with the smallest distance change are selected as the optimal solution from the individuals with minimal difference in fitness values. The calculation of distance change ∆D(j) is shown in Equation (9). The nodes and their adjacency vertices in the local subgroups of road network G are adjusted on the basis of the optimal scheme. In this process, the optimal scheme may cause topological errors. However, these topological errors will be marked in the result: where i is the ith node or the adjacency vertex in G, j is its position, and ∆D is the sum of distance changes of the nodes and their adjacency vertices in G. D i is the position of the ith node or the adjacency vertex in G under seven-digit precision and D' i is the position of the ith node or the adjacency vertex in G under five-digit precision.

Experiments Design and Study Area
The proposed algorithm was implemented in C#, and the software developed for it can produce a road network map with five-digit precision from one with seven-digit precision. The seven-digit road network data (Advanced Driving Assistance System (ADAS) in specific formats) of a region in Chongqing, China, which was provided by NavInfo Co., Ltd. (Beijing, China), which included 53,525 roads, 106,537 endpoints, 8619 intersection points, and 118,327 vertices, were selected as experimental data. The study area is shown in Figure 15.
with special nodes was set to 3°. The parameters and threshold setting of the genetic algorithm are shown in Table 1. In Figure 15, the region in box 1 is shown in Figure 16, and the region in box 2 is shown in Figure 17. These values are not fixed but are deemed relatively reasonable for this experiment.   Figure 16, and the region in red box 2 is shown in Figure 17.

Result Analysis
The improved Opheim algorithm proposed in this study was compared with the classical DP, Visvalingam-Whyatt (VW), and Opheim algorithms. The data compression rates of the four simplification methods were similar, facilitating comparison of the simplification effect. The vertical distance threshold was set to 0.5 m in the DP algorithm, the area threshold was set to 4 m 2 in the VW algorithm, and the minimum distance constraint was set to 5 m in the Opheim algorithm. The simplified results after maintenance are shown in Figure 16. From the visual effect of the simplified line, the distance among points best met the distance constraint and showed a good approximation of the original road. Compared with our algorithm, the DP algorithm produced a better compression result for straight, long, and flat road sections, and the compression effect of the curved roads was  Figure 16, and the region in red box 2 is shown in Figure 17. not ideal. Especially for curved roads with large bends and dense points, the distance constraint was difficult to meet. The regions outlined by the black dotted lines in Figure 16b show the simplified version using the DP and the proposed algorithms with five-digit precision. A considerable difference was found in whether the local road shape was maintained. The results showed that the simplified result of our algorithm is was better than that of the DP algorithm.
The simplified result of the VW algorithm is shown in Figure 16c. A few points do not satisfy the distance constraint. The road shape under five-digit precision changed more compared with the original shape, as shown in the dotted box, because road shape maintenance was disregarded.
The Opheim algorithm is similar to the current algorithm, but it does not consider the changes in turning angle and road shape. As a result, the road shape considerably changed after reducing precision. The simplification result was also not as good as that of the proposed algorithm.
(a)  Figure 17 shows four simplified results for a ring road in region 2 in Figure 15. Except for the improved Opheim algorithm proposed in this study, the DP, VW, and Opheim algorithms were unable to maintain the spatial topological consistent with the original road. The topological errors reserved points. Figure 18 is a result of maintaining the spatial directional relationship of the road network associated with the special nodes in Figure 13 by using e-GA. This result of the intelligent algorithm is generally optimal. This method keeps the morphological characteristics and displacement precision of the road network. In Figure 18, the overall azimuth deviation is the smallest, and the similarity is the highest. Given the requirements of road network data with five-digit precision, the distance constraint of vertex elimination was 2 m. The compression rate is related to the density of points in the original roads. The distance constraint ε in this experiment was set to 5 m to display the simplification details. When recognizing long straight lines, the bandwidth µ of the strip was 0.1 m, the length threshold γ of the long straight line was 100 m, and the change threshold of turning angle was 5 • in the process of local shape maintenance. In the process of maintaining the spatial directional relationship of special nodes, the change threshold η 1 of the azimuth of each road segment associated with special nodes was set to 3 • . The parameters and threshold setting of the genetic algorithm are shown in Table 1. In Figure 15, the region in box 1 is shown in Figure 16, and the region in box 2 is shown in Figure 17. These values are not fixed but are deemed relatively reasonable for this experiment.

Result Analysis
The improved Opheim algorithm proposed in this study was compared with the classical DP, Visvalingam-Whyatt (VW), and Opheim algorithms. The data compression rates of the four simplification methods were similar, facilitating comparison of the simplification effect. The vertical distance threshold was set to 0.5 m in the DP algorithm, the area threshold was set to 4 m 2 in the VW algorithm, and the minimum distance constraint was set to 5 m in the Opheim algorithm. The simplified results after maintenance are shown in Figure 16. From the visual effect of the simplified line, the distance among points best met the distance constraint and showed a good approximation of the original road. Compared with our algorithm, the DP algorithm produced a better compression result for straight, long, and flat road sections, and the compression effect of the curved roads was not ideal. Especially for curved roads with large bends and dense points, the distance constraint was difficult to meet.
The regions outlined by the black dotted lines in Figure 16b show the simplified version using the DP and the proposed algorithms with five-digit precision. A considerable difference was found in whether the local road shape was maintained. The results showed that the simplified result of our algorithm is was better than that of the DP algorithm.
The simplified result of the VW algorithm is shown in Figure 16c. A few points do not satisfy the distance constraint. The road shape under five-digit precision changed more compared with the original shape, as shown in the dotted box, because road shape maintenance was disregarded.
The Opheim algorithm is similar to the current algorithm, but it does not consider the changes in turning angle and road shape. As a result, the road shape considerably changed after reducing precision. The simplification result was also not as good as that of the proposed algorithm. Figure 17 shows four simplified results for a ring road in region 2 in Figure 15. Except for the improved Opheim algorithm proposed in this study, the DP, VW, and Opheim algorithms were unable to maintain the spatial topological consistent with the original road. The topological errors are presented in the dotted box in Figure 17. The result of the improved Opheim algorithm, shown in Figure 17a, implied that eliminating point P will lose topological consistency after simplification. Accordingly, point P was preserved and marked as a problem point. Vertex elimination under high precision guarantees results without self-intersections, regardless of the morphology of the input line and the tolerance value. In contrast, the DP, VW, and Opheim algorithms eliminated important characteristic points and violated the positional precision constraint. Figure 17b-d depict that some intersection points on the road are eliminated. The improved Opheim algorithm retained the intersection points and other feature points while satisfying the distance constraint. Another advantage of the proposed method shown in Figure 17 is that it can meet the distance requirement of road data compression better than the other methods, which do not consider the distance among reserved points. Figure 18 is a result of maintaining the spatial directional relationship of the road network associated with the special nodes in Figure 13 by using e-GA. This result of the intelligent algorithm is generally optimal. This method keeps the morphological characteristics and displacement precision of the road network. In Figure 18, the overall azimuth deviation is the smallest, and the similarity is the highest.
To better demonstrate the performance of our method, experimental results based on the data from Chengdu (61,641 roads), Guangzhou (71,326 roads), and Shanghai (43,571 roads) in China were compared and analyzed. Table 2 presents the comparisons of the four groups of experiments. Similarly, the compression rates of these four sets of experiments were comparable. The assessment results of the modified Hausdorff distance (MHD) [49] between the original and the simplified roads are shown in Table 2, which shows a higher shape similarity degree with our proposed method than the other three methods. Table 3 shows the changes of the azimuth angles with and without adjustment of the spatial directional relationship. Table 4 describes the changes in the turning angles with and without adjustment of the spatial directional relationship. Our method was the key to extracting road network maps with five-digit precision from those with seven-digit precision.  To better demonstrate the performance of our method, experimental results based on the data from Chengdu (61,641 roads), Guangzhou (71,326 roads), and Shanghai (43,571 roads) in China were compared and analyzed. Table 2 presents the comparisons of the four groups of experiments. Similarly, the compression rates of these four sets of experiments were comparable. The assessment results of the modified Hausdorff distance (MHD) [49] between the original and the simplified roads are shown in Table 2, which shows a higher shape similarity degree with our proposed method than the other three methods. Table 3 shows the changes of the azimuth angles with and without adjustment of the spatial directional relationship. Table 4 describes the changes in the turning angles with and without adjustment of the spatial directional relationship. Our method was the key to extracting road network maps with five-digit precision from those with seven-digit precision.  Table 3. The average changes in the azimuth angle in the spatial directional relationship.  Table 4. The average changes in the turning angle in the spatial directional relationship.

Changes in Turning Angle
Chongqing Chengdu Guangzhou Shanghai

Conclusions
In this study, we innovatively proposed a method that combines the improved Opheim algorithm with the intelligent maintenance of the spatial directional relationship. Based on the grid model, this new method effectively expresses the process of transforming road network maps from seven-to five-digit precision, which involves four main steps. First, the improved Opheim algorithm under seven-digit precision deletes vertices that do not meet requirements for navigation. Second, local shape maintenance under five-digit precision is used to maintain the spatial directional relationship between the road segments connecting with the vertices. Third, the spatial directional relationship at a general node is adjusted. Finally, we explained the problems caused by the method when maintaining the spatial direction relationship between roads at special nodes, and introduced the process of solving this problem with a genetic algorithm with the elite strategy in detail. To illustrate the application of our method, we examined the road network extraction in four cities, and we proved these results to be reasonable and effective. The threshold in the experiment was set on the basis of experience. The following conclusions can be drawn from the results: (1) Compared with the DP, VW, and Opheim algorithms, the improved Opheim algorithm that considers requirements in navigation retains the characteristic points needed for navigation and is combined with the intelligent maintenance of the spatial directional relationship. The local shape of the roads is similar to the original roads after road data precision is reduced. In the Hausdorff distance statistics between the road networks with five-digit precision and the original road networks, the result of our method was close to the Opheim algorithm, and the difference was smaller than those of the other three methods. The average change of turning angles with local shape adjustment was optimized compared to the average change in turning angles without local shape adjustment. A part of the visual effects of the results was shown in Figure 16. (2) Our algorithm considers the spatial topological relationships among all features. Theoretically, in the process of maintaining the local shape of the road, when topological errors occur in all optional positions of a vertex, topological errors will be marked in the result. During the process of adjusting the spatial relationship among the roads connecting with the node, if all the schemes satisfying the user-defined constraint on spatial directional deviation cause spatial topological errors, the optimal scheme selected from them will also cause spatial topological errors. (3) The spatial directional relationships between roads at all nodes in road networks are maintained when data precision is lowered. The spatial directional relationship among road segments connected at every general node is adjusted optimally and total changes in azimuths of road segments connected at special nodes are minimized by obtaining the near-optimal solution.
However, the proposed method in this paper was used to extract road network maps (five digit) used for conventional navigation from high-precision road network maps (seven digit) instead of extracting maps with other precision levels. If it will be used to produce road network data with four-digit precision, then the grid width is expanded by 10 times, and the simplification and adjustment of a road network will be more difficult. In future research, we hope that this proposed algorithm becomes more intelligent to meet the requirements of road network generalization for navigation. Another future task is to study the iterative displacement method to address the spatial topological conflicts caused by road simplification, and to consider the buildings along roads to avoid inconsistencies between roads and buildings after reducing the precision.