In recent years, the ability of novice contributors to accurately describe spatial geographic data poses many concerns for OSM data quality [8]. Therefore, many studies have focused on evaluating the accuracy and completeness of OSM data. International Cartographic Association (ICA) [28] developed seven rules for assessing the quality of spatial geographic data. Based on these seven rules [29], Barron et al. extended the rules, including the semantic, geometric accuracy, and availability of spatial data. Moreover, since contributors are an important part of map production, data-centric and contributor-centric assessments are often combined [11,12,30,31,32,33]. Overall, the OSM quality assessment can be divided into extrinsic quality measures and intrinsic quality measures.

Extrinsic quality measures: This type of research compares OSM data with authoritative data from other official institutions for external evaluation [10,34,35,36]. Such methods rely on external data. However, authoritative datasets are more difficult to obtain than public ones, and the update efficiency of their data is sometimes slower than the fast-changing OSM data. All these factors constrain such evaluation methods in applications and expansions.

Intrinsic quality measures: Intrinsic method does not rely on external or authoritative data sources for validation [37,38]. The methods can measure the accuracy of OSM data by assessing changes in historical versions of the data, or by associating user reputation. As an example, Fogliaroni et al. calculated the quality score of geographic features by analyzing the geometric, qualitative, and semantic changes in the edited version of history. They used it to approximate the quality score of spatial data at the end [8]. Zhou and Zhao used spatial similarity and geometrical similarity to calculate the similarities between the versions. The reputation of the contributor was obtained by analyzing the complicit assessments computed by version similarity [28]. Mullagann et al. analyzed the spatial semantic relations of point features. The spatial semantic interaction was used to measure the semantic similarity of the change history of geographic elements [33].

The quality improvement of OSM data has received much attention from many researchers. It can be divided into two aspects: identifying and correcting erroneous data for OSM, and OSM tag recommendation.

Identifying and correcting erroneous data for OSM: The early literature on improving the quality of crowd-sourced geographic data focused on detecting and modifying error elements. For example, Vargas et al. used a Markov random field method to maximize the correlation among annotations of OSM buildings and predicted building probability maps. After removing several redundant geometric annotations through the relationship between building probability maps and thresholds, they used CNN to predict and add new architectural geometric annotations [7]. Kashian et al. analyzed the “semantics” of the newly contributed data by identifying potential patterns of coexistence between POIs and other geographic features. They calculated the likelihood of a POI being registered at this location to improve the detection and verification system of the OSM platform. The location accuracy of registered POI in OSM can be improved [13]. These studies have contributed significantly to improving the semantic quality of OSM databases.

OSM tag recommendation: OSM does not have a proper tag verification mechanism, which leads to a problem in that the OSM tags vary greatly with different contributors. A tag recommendation method is a good method for this issue and can significantly improve the quality of OSM data [8]. There have been many studies on OSM and other open data in recent years. For example, Arnaud et al. established a tag recommendation system named “OSMantic”. By calculating the corresponding semantic similarity score, the system, which gives timely relevant suggestion tags by calculating the corresponding semantic similarity score when users submit commit them. And the system will give some semantic accuracy hints when the score is too low [14]. Storandt et al. developed a recommendation system that only needs the POI name to recommend appropriate tags [15]. Jilani et al. constructed the features of road elements, such as degree distribution, intermediary centrality, node number in a bounding box, etc. The model constructs and represents the road and its features by using a graph structure. They used an artificial neural network (ANN) to train the model and recommend tags [16]. Corcoran et al. focus on geometric features and define a series of geometric features about road elements, such as degree, road curvature, parallelism, etc. Finally, their model reported 68% and 65% weighted accuracy and recall values, respectively [18]. Hacar used geometric and semantic features to classify and recommend the leisure tags of polygons [19]. Tag recommendation can motivate contributors to contribute correct tags which are highly effective in improving quality. Therefore, it is currently a widely studied method. However, these studies only consider geometric or other semantic features of the element itself, which mostly depend on the quality of the OSM data, and it often struggles to adapt to semantically incomplete datasets.

For geographical spatial features, each element is related to other similar elements, and only the distance determines the size of the influence [27]. In the current research, there is a tendency to combine features of geographic elements with the spatial context, and there are many achievements. For example, Zhang et al. used geometric features and restricted features of road elements to detect label tag semantic inconsistency and other problems, while giving intelligent suggestions based on the information available in the spatial context of the problem data [20]. Alghanim et al. used building context as a feature to analyze road elements and used a 20 M to 200 M linear buffer to count context semantics. They developed classifiers to recommend classification tags for road elements based on this method [9]. Ali et al. identified and predicted several grassland fuzzy categories based on contextual attributes and topological features by analyzing the case of building elements and road elements versus object elements with three selected topologies. Among them, the connecting road context applies several road element categories related to park grass, including “foot” and “bike” [21].