Road network selection is critical in map generalization. While machine learning-based models improve selection performance, their increased complexity reduces explainability, which complicates the explicit description of the relationship between input road features and selection decisions. To address this issue, we propose an explainable artificial intelligence (XAI) approach for road network selection that treats road strokes as processing units and extracts the descriptive features for each stroke. A neural additive model (NAM) that consists of several independent feature networks was used to analyze descriptive features to determine whether each stroke should be retained. The explainability of the XAI approach was driven by the learning of linearly combined feature networks in the NAM, which clearly differentiates the contribution of each stroke’s features to the final selection results. To balance explainability and performance, knowledge distillation was employed to transfer knowledge from a teacher model to the NAM student model. Experiments on datasets showed that the XAI approach achieved over 87% consistency with manual selection. Notably, it provided a mechanism for both global and local explainability analyses, thereby improving our understanding of why certain strokes are retained or deleted. These insights into the model’s decision-making process help advance the automation of map generalization.

Understanding individual building functions is crucial for analyzing broader urban dynamics. Existing studies have incorporated multi-source data to predict building functions but often neglect the spatial and semantic relations between buildings and their surrounding geographic entities. This study proposes a knowledge graph-based approach for geographic context-aware building function prediction to address this gap. Based on multi-source data, we first constructed a building knowledge graph (BuildingKG), where nodes represent buildings and other related entities, and edges denote their spatial and semantic relations. Subsequently, we applied a knowledge graph embedding technique to the BuildingKG to capture the structural knowledge of entities, which encodes the interactions between them. The structural knowledge, combined with the attributes of individual entities, was concatenated to form the descriptive features of the entity nodes. Finally, we designed a meta-path-based graph transformer neural network model to process the graph, comprising nodes and edges from the BuildingKG, to classify building function types using a supervised learning method. Experiments demonstrated that our approach achieved an overall accuracy of 93.05%, outperforming baseline models with superior computational efficiency. Moreover, interpretability analysis revealed different entity features’ global and local contributions to the prediction results, underscoring the importance of considering geographic context.

Landform type recognition presents significant implications for understanding landform origins, evolutionary mechanisms, and morphological differences. Artificial intelligence (AI) techniques based on sample learning often lead to unsatisfactory outcomes due to the intricate genesis and regional heterogeneity of landforms. This study combines domain knowledge with a deep learning (DL) model to improve landform type recognition. Contour data serves as a valuable resource, offering rich morphological information across horizontal, vertical, local, and macro scales. Our approach incorporated morphological knowledge and proximity relationships derived from contours into a graph convolutional network using the DiffPool technique (GCN-DP). Guided by the First Law of Geography, contours within each landform unit were represented as graphs, incorporating morphological knowledge as node features. The GCN-DP model then employed convolution and pooling to extract hierarchical features from these graphs for landform type recognition. A performance evaluation demonstrated the effectiveness of our method with an F1-score of 87.40%, surpassing RF and GCN methods by 5.24–12.50%, respectively. Ablation experiments confirmed the usefulness of morphological knowledge. This study offers an efficient strategy for landform type recognition, improving the level of intelligent mining using contour data.

Points of interest (POIs) are critical components of maps, often accompanied by annotations that convey essential information. However, when maps are scaled down, annotation sizes generally remain unchanged, frequently resulting in overlaps and conflicts that compromise map clarity. To address this issue, this article incorporates annotation conflicts as constraints within the POI selection decision-making process during scale transformations. It presents a novel POI selection method using a graph convolutional network (GCN) considering annotation conflicts. The proposed method begins by constructing a graph structure based on Delaunay triangulation, which represents second-order proximity relationships. Node features are then extracted from three dimensions: semantic, spatial and annotation, while annotation conflicts are abstracted as edge weights within the graph. Leveraging the TAGCN network, a POI selection model is developed, enabling intelligent POI selection through semi-supervised training. This approach transforms the POI selection task into a classification problem, seamlessly integrating expert knowledge into the deep learning framework through samples and models. Moreover, this article introduces an annotation placement algorithm to further mitigate annotation conflicts. Experimental results demonstrate that this method effectively reduces annotation conflicts while preserving map details, outperforming Maplex in ArcGIS.

Using multi-modal data to learn region representations has gained popularity for its ability to reveal diverse socioeconomic features in cities. However, many studies focus solely on semantic features from points-of-interest (POIs), neglecting the issue of spatial imbalance. This article introduces a Multi-Graph Representation Learning framework for Region Embedding (MGRL4RE), which leverages both inter-region and intra-region correlations through two main components: multi-graph construction based on various region correlations and multi-graph representation learning. The construction module creates a multi-graph reflecting various correlations among regions, utilizing geo-tagged POIs, region data, and human mobility data. Specifically, we assess a region’s importance relative to its spatial context (neighborhood) and develop spatially invariant semantic features to address spatial imbalance. Furthermore, the representation learning module generates comprehensive and effective region representations via multi-view embedding fusion. Our extensive experiments across various downstream tasks, including land use clustering, region popularity prediction, and crime prediction, confirm that our model significantly outperforms existing state-of-the-art region embedding methods.

Fine-grained urban flow inference (FUFI) is crucial for traffic management, as it infers high-resolution urban flow maps from coarse-grained observations. Existing FUFI methods typically focus on a single city and rely on comprehensive training with large-scale datasets to achieve precise inferences. However, data availability in developing cities may be limited, posing challenges to the development of well-performing models. To address this issue, we propose cross-city fine-grained urban flow inference, which aims to transfer spatio-temporal knowledge from data-rich cities to data-scarce areas using meta-transfer learning. This paper devises a Spatio-Temporal Deviation Alignment (STDA) framework to mitigate spatio-temporal distribution deviations and urban structural deviations between multiple source cities and the target city. Furthermore, STDA presents a cross-city normalization method that adaptively combines batch and instance normalization to maintain consistency between city-variant and city-invariant features. Besides, we design an urban structure alignment module to align spatial topological differences across cities. STDA effectively reduces distribution and structural deviations among different datasets while avoiding negative transfer. Extensive experiments conducted on three real-world datasets demonstrate that STDA consistently outperforms state-of-the-art baselines.

Numerous methods based on square rasters have been proposed for polygon generalization. However, these methods ignore the inconsistent distance measurement among neighborhoods of squares, which may result in an imbalanced generalization in different directions. As an alternative raster, a hexagon has consistent connectivity and isotropic neighborhoods. This study proposed a hexagon-based method for polygon generalization using morphological operators. First, we defined three generalization operators: aggregation, elimination, and line simplification, based on hexagonal morphological operations. We then used corrective operations with selection, skeleton, and exaggeration to detect, classify, and correct the unreasonably reduced narrow parts of the polygons. To assess the effectiveness of the proposed method, we conducted experiments comparing the hexagonal raster to square raster and vector data. Unlike vector-based methods in which various algorithms simplified either areal objects or exterior boundaries, the hexagon-based method performed both simplifications simultaneously. Compared to the square-based method, the results of the hexagon-based method were more balanced in all neighborhood directions, matched better with the original polygons, and had smoother simplified boundaries. Moreover, it performed with shorter running time than the square-based method, where the minimal time difference was less than 1 min, and the maximal time difference reached more than 50 mins.

The past decade has witnessed the increasingly created point-of-interest (POI) data, which are utilized to express the semantics of places. To understand the POI semantics, current methods usually embed POI categories into a latent space via certain trajectory sequential models, while neglecting the underlining spatial information. It is noteworthy that the complex spatial relationships among POI categories contain substantial information that benefits meaningful semantic embeddings for various POI categories. Inspired by this, we present a unified POI Category Embedding Method (CatEM for short), which jointly encodes the sequential transitions and spatial relations of POI categories as well as the adaptive semantic neighbors of each POI category. The merits of CatEM lie in that: (1) it considers the pairwise spatial similarities between categories and represents categories with larger similarity values with adjacent embeddings in the latent space, and (2) it adaptively locates neighbor categories with similar semantics in the embedding space to improve the adaptivity of POI category embedding. The proposed pre-trained POI category embeddings are justified by three downstream tasks. Extensive experiments demonstrate the superiority of our proposed model, as compared to several cutting-edge baselines.

Detecting interchanges in road networks benefit many applications, such as vehicle navigation and map generalization. Traditional approaches use manually defined rules based on geometric, topological, or both properties, and thus can present challenges for structurally complex interchange. To overcome this drawback, we propose a graph-based deep learning approach for interchange detection. First, we model the road network as a graph in which the nodes represent road segments, and the edges represent their connections. The proposed approach computes the shape measures and contextual properties of individual road segments for features characterizing the associated nodes in the graph. Next, a semi-supervised approach uses these features and limited labeled interchanges to train a graph convolutional network that classifies these road segments into an interchange and non-interchange segments. Finally, an adaptive clustering approach groups the detected interchange segments into interchanges. Our experiment with the road networks of Beijing and Wuhan achieved a classification accuracy >95% at a label rate of 10%. Moreover, the interchange detection precision and recall were 79.6 and 75.7% on the Beijing dataset and 80.6 and 74.8% on the Wuhan dataset, respectively, which were 18.3–36.1 and 17.4–19.4% higher than those of the existing approaches based on characteristic node clustering.

Research has developed numerous algorithms to simplify building data. Each algorithm has strengths and weaknesses in addressing shape characteristics, but no single algorithm can appropriately simplify all buildings. This study proposes a hybrid approach that identifies the best simplified representation of a building among four existing algorithms. The proposed approach applies the four algorithms to generate simplification candidates. With a backpropagation neural network, an evaluator is built through supervised learning based on measurements describing the changes in position, size, orientation, and shape between the original building and the candidates of its simplified representations. The evaluator determines the most appropriate candidate. Experiments on buildings from residential and commercial areas in Shenzhen city show that the hybrid approach can combine the advantages of different algorithms. The percentages of unreasonable simplified buildings in the results obtained using the hybrid algorithm are 3.8% in the residential area and 0 in the commercial area, respectively, which are significantly lower than those in the results of standalone applications of the four algorithms. Furthermore, comparison with the simplification algorithm in the popular software, ArcGIS, confirms that our approach shows better results in terms of corner squaring and maintaining the regional characteristics of buildings.

The shape of a geospatial object is an important characteristic and a significant factor in spatial cognition. Existing shape representation methods for vector-structured objects in the map space are mainly based on geometric and statistical measures. Considering that shape is complicated and cognitively related, this study develops a learning strategy to combine multiple features extracted from its boundary and obtain a reasonable shape representation. Taking building data as example, this study first models the shape of a building using a graph structure and extracts multiple features for each vertex based on the local and regional structures. A graph convolutional autoencoder (GCAE) model comprising graph convolution and autoencoder architecture is proposed to analyze the modeled graph and realize shape coding through unsupervised learning. Experiments show that the GCAE model can produce a cognitively compliant shape coding, with the ability to distinguish different shapes. It outperforms existing methods in terms of similarity measurements. Furthermore, the shape coding is experimentally proven to be effective in representing the local and global characteristics of building shape in application scenarios such as shape retrieval and matching.

As a basic and significant operator in map generalization, polyline simplification needs to work across scales. Perkal’s ε-circle rolling approach, in which a circle with diameter ε is rolled on both sides of the polyline so that the small bend features can be detected and removed, is considered as one of the few scale-driven solutions. However, the envelope computation, which is a key part of this method, has been difficult to implement. Here, we present a computational method that implements Perkal’s proposal. To simulate the effects of a rolling circle, Delaunay triangulation is used to detect bend features and further to construct the envelope structure around a polyline. Then, different connection methods within the enveloping area are provided to output the abstracted result, and a strategy to determine the best connection method is explored. Experiments with real land-use polygon data are implemented, and comparison with other algorithms is discussed. In addition to the scale-specificity inherited from Perkal’s proposal, the results show that the proposed algorithm can preserve the main shape of the polyline and meet the area-maintaining constraint during large-scale change. This algorithm is also free from self-intersection.

In map generalization, the displacement operation attempts to resolve proximity conflicts to guarantee map legibility. Owing to the limited representation space, conflicts may occur between both the same and different features under different contexts. A successful displacement should settle multiple conflicts, suppress the generation of secondary conflicts after moving some objects, and preserve the distribution patterns. The effect of displacement can be understood as a force that pushes related objects away with properties of propagation and distance decay. This study borrows the idea of vector fields from physics discipline and establishes a vector field model to handle the displacement of multiple conflicts in building generalization. A scalar field is first constructed based on a Delaunay triangulation skeleton to partition the buildings being examined (e.g., a street block). Then, we build a vector field to conduct displacement measurements through the detection of conflicts from multiple sources. The direction and magnitude of the displacement force are computed based on an iso-line model of vector field. The experiment shows that this global method can settle multiple conflicts and preserve the spatial relations and important building patterns.