Continued Advances in AI-Driven Geospatial Analysis and Data Generation

1. Introduction

GeoAI has maintained strong momentum since the conclusion of the 2024 Special Issue on “Advances in AI-Driven Geospatial Analytics and Data Generation” [1]. While journal special issues remain an important method for disseminating the latest advances, the GeoAI research landscape is also expanding through dedicated conferences, symposia, and other scholarly venues that provide additional opportunities for exchange and visibility. For example, June 2026 marked both the inaugural International GeoAI Conference in Ghent, Belgium (https://geoaiconference.org) and the University Consortium for Geographic Information Science (UCGIS) annual symposium, themed “AI Everywhere” (https://www.ucgis.org/site/symposium-2026, accessed on 23 August 2026), underscoring the community’s sustained interest in AI applications in geospatial contexts. In line with this momentum, and building on the success of the first edition, this Editorial introduces the Special Issue “Advances in AI-Driven Geospatial Analysis and Data Generation (2nd Edition)” (https://www.mdpi.com/journal/ijgi/special_issues/G2VCJ097OM, accessed on 23 August 2026), which aims to showcase new methodologies and emerging applications.

Recent systematic reviews highlight a substantial surge in GeoAI literature, driven by the rapid integration of generative AI (GenAI) and large language models (LLMs) into spatial workflows [2,3]. To structure this expanding landscape, Wang et al., using scientometric and bibliometric analyses, categorize the operational roles of GenAI into distinct functions based on information gain: purifiers for targeted data extraction and feature engineering, converters for cross-modal transformations, generators for synthetic data augmentation, and reasoners for human-like spatial problem-solving [3]. This review also notes that while text remains the predominant data input, applications are diversifying rapidly across fields like cartography, spatial analysis, informatics, and transportation, with growing interest in multimodal workflows that integrate text, imagery, maps, trajectories, and other spatio-temporal data sources.

A major focus within the recent synthesis literature is the shift toward autonomous GIS [4] and geo-foundation models (GeoFMs) [5]. Review articles note an accelerating transition from task-specific deep learning pipelines to versatile, task-agnostic architectures. This includes the deployment of intelligent agent frameworks that leverage LLMs to translate natural language into executable geoprocessing code, effectively lowering technical barriers for non-experts [2,5]. Closely related is the rise of spatially explicit retrieval-augmented and agentic systems, in which LLMs are coupled with geospatial databases, GIS tools, APIs, and external knowledge sources to support grounded querying, tool use, and multi-step spatial reasoning rather than functioning as standalone conversational models [6]. Current progress is also defined by the emergence of GeoFMs pretrained on heterogeneous spatio-temporal data [5]. These frameworks span geospatial language models, Vision Transformers for multi-sensor remote sensing, and graph foundation models designed to internalize spatial priors, topological relationships, and scale dependencies directly within neural representations [2,5].

The following section provides some editorial insights into this Special Issue (Section 2), with subsequent sections summarizing the GeoAI methods (Section 3) and applications (Section 4) featured in the published papers. In Section 5, we discuss potential future directions for GeoAI.

2. Editorial Insights

The two successive Special Issues managed by the same editorial team offer useful insight into how GeoAI trends are evolving. Bibliometric patterns remained largely stable. We also maintained a rigorous review process: of 47 submitted manuscripts, 15 were accepted for publication. In addition, several submissions were redirected to the regular journal review track because they were not a sufficiently strong topical fit for the Special Issue. Contributions to the Special Issue come from four continents. However, representation remains uneven: only one paper includes an author from Africa, and Latin America is not represented, mirroring broader patterns reported for GIScience journal publications [7]. China is the most prominent contributor, with authors involved in 12 of the 15 published papers (80%), a share even higher than that reported for GIScience more broadly (68%) [7]. The rate of cross-country collaboration remained similarly low to that of the first edition, with only two papers (13%). This share also falls well below the 30% average reported for GIScience journal articles more broadly [7]. Finally, two-thirds of the published papers include at least one female author, suggesting a shift in gender representation against the backdrop of the broader underrepresentation of women in authorship and their historically lower manuscript acceptance rates in related quantitative fields [8].

3. Applied GeoAI Methods

The papers in this Special Issue employ a wide range of GeoAI methods. This section categorizes and describes the methods used and identifies methodological shifts between the 2024 and 2026 Special Issues.

3.1. GeoAI Methods in the Current Special Issue

The papers in this Special Issue can be organized into three methodological clusters, which are also listed in Table 1.

Table 1. Special Issue contributions by methodological cluster.

Paper (Ref)	Core GeoAI/AI Architecture	Geospatial Application
1. Predictive Spatio-Temporal GeoAI
Li et al.	UPTRec (User graph + POI-transition graph + Transformer)	Next-POI recommendation
Zhou et al.	TAGAT-LSTM-trans (Graph Attention + LSTM + Transformer)	Traffic flow forecasting
Lee and Kang	PGTFT (Graph Attention + Temporal Fusion Transformer)	Pedestrian congestion prediction
Xiang et al.	TPE-XGB (AutoML + Extreme Gradient Boosting)	Soil nutrient prediction and mapping
Wang and Hu	GLEN (GCN + LSTM ensemble network)	Urban multimodal traffic flow prediction
Yu et al. (Trajectory/Social)	TraFriendFormer (Graph Neural Network + Transformer)	Next-POI recommendation
Brahimi et al.	eX-STIN (Spatio-temporal inference network + explainable AI)	Car-sharing demand prediction
Gao et al.	Dynamic STGCN (Grid partition + graph convolution)	Large-scale traffic flow forecasting
2. Representation Learning & Structure
Lun et al.	ReMVL-Net (Multi-view representation learning)	Next-location prediction
Kleebauer et al.	RetinaNet (FPN + focal loss)	Wind turbine mapping and coordinate correction
Hua et al.	Contrastive representation learning (Image deformation + NT-Xent loss)	Urban morphology discovery
Yang et al.	Graph representation learning (iGPN + PR propagation)	Functionally cohesive building group recognition
3. Generative & Interface-Oriented GeoAI
Wang and Li	Stacking ensemble learning (PSPNet + RF/XGBoost/LightGBM)	Street visual quality evaluation
Yu et al. (SceneDiffusion)	SceneDiffusion (Attention-aware Graph + Layout Conditional Diffusion)	Geographically coherent scene generation
Sherman et al.	Fine-tuned ChatGPT-4o-mini	Natural language-based geospatial analysis

• The first cluster, with eight papers, focuses on predictive spatio-temporal GeoAI, especially for movement, traffic, and demand forecasting, and is primarily concerned with learning complex dependencies rather than generating new content.
• The second cluster, with four papers, emphasizes representation learning and spatial structure discovery, where the goal is to learn more meaningful geographic embeddings, communities, or morphology-aware similarity spaces from heterogeneous data.
• The third cluster, with three papers, is centered on generative and interface-oriented GeoAI, including constrained scene generation, language model-based geospatial querying, and data generation or refinement through computer vision.

In the first cluster, the dominant methodological pattern is the replacement of simple spatial adjacency and single-stream temporal modeling with richer relational architectures that jointly encode multiple forms of geographic interaction. These studies commonly build several complementary graphs or views, such as user similarity, point-of-interest transitions, road connectivity, transfer probability, functional similarity, or local grid partitions, and fuse them with temporal encoders such as long short-term memory (LSTM) networks, Transformers, gated temporal modules, or temporal fusion frameworks. A representative example includes TAGAT-LSTM-trans, a Transfer-aware Spatio-Temporal Graph Attention Network with LSTM and Transformer components, developed for traffic flow forecasting, which combines transfer probabilities and distance decay with graph attention and sequential modeling.

In the second cluster, the emphasis shifts from forecasting to learning stronger latent spatial structure through multi-view learning, contrastive representation learning, graph-based community detection, and morphology-aware similarity modeling, with the common idea that geographic proximity should be enriched with semantic, functional, or relational context. Examples include representation-enhanced multi-view learning networks for predicting the next location of urban individuals, which combine community-aware spatial encoding with multi-granular temporal representations, as well as refined contrastive learning frameworks for urban morphology discovery, which adapt the contrastive loss function so that it better reflects similarities in urban form.

In the third cluster, GeoAI is extended beyond prediction toward generation, interaction, and data curation through spatially constrained diffusion models that produce geographically coherent scenes, deep learning workflows that refine mapped object locations and attributes, and large language model pipelines that translate natural language requests into geospatial computation. A generative AI example in this cluster is SceneDiffusion, which combines a Spatial relationship Attention-aware Graph module with a Layout information-guided Conditional Diffusion module to generate scenes that satisfy topological, directional, and distance constraints.

The papers span a spectrum from strongly spatial GeoAI to more general artificial intelligence methods applied to geospatial problems. At one end, several studies are clearly GeoAI in a methodological sense because they explicitly model geographic structure, such as spatial relationships, network dependence, topology, distance, directional constraints, or multi-scale spatio-temporal interaction. At the other end, some papers use more general machine learning, computer vision, ensemble learning, or language model approaches on spatial datasets without deeply adapting the core method to geographic reasoning. Overall, the 2026 Special Issue leans clearly toward true GeoAI, but it also includes a meaningful share of studies that are better described as artificial intelligence for geospatial analysis rather than inherently spatial AI methods.

3.2. Change in GeoAI Methods Between 2024 and 2026 Special Issues

There are a few notable methodological shifts between the two Special Issues. The 2024 collection is more heterogeneous and exploratory, with a mix of classical spatial models hybridized with machine learning, convolutional and recurrent neural networks, semi-supervised learning, generative adversarial networks, early Transformer use, and a first-wave of large language or vision–language model experiments. In that collection, many solution approaches still rely on rasterized spatial representations, image-like encodings, or general-purpose deep learning architectures adapted to spatial data. By contrast, the 2026 collection is more concentrated around explicitly relational and composite GeoAI pipelines where graph-based spatio-temporal learning becomes much more central, alongside broader shifts toward multi-view and multi-source fusion, stronger representation learning, increased use of explainability modules, and the emergence of generative and language model-based geospatial workflows. In its papers, some previously used methods seem to have been absorbed into more recent frameworks. For example, convolutional LSTM models, basic graph convolution, standalone ensemble learning, and generic boosting remain useful but appear more often embedded within richer multi-component systems. Notable advances are the stronger use of multi-graph fusion and multi-view fusion [9,10], a shift from sequential temporal modeling toward hybrid temporal encoding capturing periodicity, long-range dependencies, event sensitivity, and cross-scale temporal structure [11], and a move from single-model architectures toward modular composite frameworks [12], some of which combine graph learning, attention mechanisms, sequence encoders, explainability components, and generative modules within one pipeline.

Overall, it can be observed that earlier general deep learning and hybrid approaches evolved recently into more modular, spatially aware, and representation-rich GeoAI systems.

4. Applications of GeoAI in the Special Issue

The papers published in this Special Issue clearly demonstrate how AI is actively and diversely reshaping geospatial analysis. Rather than simply applying advanced AI algorithms to geospatial datasets, authors are developing domain-specific GeoAI models and algorithms to solve diverse spatial problems. As a result, four distinct application topics emerge, ranging from urban mobility modeling to environment monitoring using generative AI. Table 1 lists the application areas together with the GeoAI methods used in the papers.

4.1. Urban Transportation and Traffic Prediction

Modeling and predicting human mobility patterns across cities remains a core challenge in urban analytics, and several papers in this Special Issue address it using advanced spatio-temporal neural networks, including Graph Neural Networks (GNNs) and Transformers. While these studies share a common methodological foundation in spatio-temporal deep learning, they differ substantially in scale, network design, and transportation context. For example, Zhou et al. and Gao et al. both model macro-scale urban vehicle traffic, but pursue different modeling strategies. Zhou et al. emphasize transfer-aware attention mechanisms to better capture dynamic spatial dependencies, whereas Gao et al. use grid partitioning to improve scalability across large spatial domains. Wang and Hu extend this line of work further by developing an ensemble framework specifically designed for multimodal traffic prediction.

In contrast to vehicular traffic studies, Lee and Kang focus on urban micro-mobility by predicting pedestrian congestion in localized urban shadow areas where direct sensing is limited. Finally, Brahimi et al. address a persistent challenge in GeoAI-based urban mobility research—model explainability—by developing an interpretable spatio-temporal inference framework for car-sharing demand prediction.

4.2. Human Behavior and Social Dynamics

The goal of predicting where a person will go next is central to modern location-based services (LBSs). Three papers in this cluster address this problem through a shared strategy of learning multi-view latent representations that capture complex mobility behavior and social context, but they differ in the types of contextual information they emphasize. Li et al. build their UPTRec model primarily around the sequences of Points-of-Interest (POIs) and temporal patterns. While Lun et al. take a similar multi-view approach but aim at broader urban localization, Yu et al. place stronger emphasis on social influence through a learning-based framework that combines Graph Neural Networks and Transformers, showing how friendship relations and movement trajectories can be modeled jointly to improve next-POI recommendation.

4.3. Urban Morphology, Visual Quality, and Built Environment

Modern AI, with its strong multimodal capabilities, has proven to be a powerful tool for capturing the physical layout of cities from diverse forms of urban big data. The papers in this direction extract structural and urban perceptions from complex, unstructured data, though they target entirely different facets of the urban built environment. For instance, Wang and Li use multi-model data fusion to assess the subjective visual quality of urban streets, whereas Hua et al. apply contrastive learning to objectively map broader urban morphology.

Taking a topological approach rather than a visual one, Yang et al. use flow-based community search to group functionally related buildings within commercial zones based on how people move between them. From a built environment perspective, Kleebauer et al. focus on the mapping of wind turbines in South Africa, demonstrating how deep learning can be used to improve the accuracy of crowdsourced vector data and analyze the spatial footprint of a specific infrastructure type in OpenStreetMap.

4.4. Generative GeoAI and Environmental Monitoring

The final group of papers focuses on generative and automated GeoAI. Yu et al. and Sherman et al. both engage with generative AI, but from markedly different angles. Yu et al. introduce SceneDiffusion, a model designed to generate visual scenes that satisfy explicit spatial constraints. Sherman et al., by contrast, focus on the analytical workflow itself by fine-tuning LLMs to translate natural language prompts into executable Python code for geospatial analysis. Lastly, extending GeoAI beyond urban environments, Xiang et al. apply automated machine learning (AutoML) to hyperspectral data to map soil nutrients, showing how advanced sensing and AI-based analysis can streamline environmental monitoring and spatial assessment.

5. Future Directions

While the field is rapidly shifting towards task-agnostic GeoFMs [5], current architectures still rely heavily on adaptations of natural language and computer vision paradigms. Future research must therefore move beyond fine-tuning generic LLMs or Vision Transformers and instead focus on developing spatially native architectures.

Agentic GeoAI and autonomous GIS represent a paradigm shift from passive, human-driven analytical tools to active, self-executing systems [4]. By leveraging LLMs as their core reasoning engine, these systems decompose high-level natural language objectives into structured workflows, including geospatial analysis [13], webGIS development [14], or remote sensing [15]. Instead of relying on manual data preparation and tool selection, intelligent geospatial agents can autonomously search for data, generate executable code, perform spatial operations through existing APIs and software libraries, and evaluate the resulting outputs. This capability transforms spatial models from static analysis pipelines into operational entities that can reason sequentially across spatial constraints, continuously interpret spatial data, and automate decision-making cycles with minimal human intervention.

However, the integration of LLMs into Agentic GeoAI introduces vulnerabilities in both spatial reasoning and value alignment. While these autonomous frameworks can translate natural language into executable code or geospatial workflows, they frequently lack a grounded understanding of physical constraints, topological rules, and geographic scale. This disconnect manifests as spatial hallucinations, that is, instances in which models generate invalid geometries, violate adjacency constraints, or otherwise distort fundamental spatial relationships [16,17]. Compounding this metric failure is the theoretical oversight of geo-alignment. Janowicz et al. [18] argue that GeoAI alignment cannot follow a one-size-fits-all paradigm, because societal norms, cultural values, and legal frameworks vary substantially across geographic contexts. As autonomous agents operate at a global scale with diminishing human oversight, deploying models without contextual geographic awareness risks producing outputs that are locally inappropriate, culturally insensitive, or even unlawful. Therefore, mitigating algorithmic bias and spatial inequity requires replacing standard natural language processing evaluation metrics with pluralistic, geographically aware alignment benchmarks. Future governance protocols are needed [14,19] that must embed learnable spatial structures and rigorous audit trails to ensure autonomous decisions are both topologically valid and geographically aligned.

Another critical future direction is addressing global representation, transferability, and spatial data equity. The bibliometric pattern (Section 2) reveals a persistent structural imbalance: 80% of the Special Issue publications involve authors from China, cross-country collaboration remains low, and the Global South is underrepresented. This geographic concentration necessitates mitigating data and algorithmic biases stemming from localized model training. Consequently, future research must focus on developing robust zero-shot, few-shot, and domain-adaptation techniques capable of transferring models trained on data-dense urban environments (such as those with rich POI networks and trajectory feeds) to data-sparse and underrepresented regions without performance decline.

More broadly, as GeoAI continues to expand rapidly and new venues for research dissemination emerge, the community must remain critical and actively address these core methodological, ethical, and structural challenges.