Large Language Models (LLMs) can be applied to various real-world tasks, many of which require spatial reasoning, like pattern-based spatial search, Point of Interest (POI) recommendation, and itinerary or travel planning. However, on their own LLMs have been shown to lack reliable spatial reasoning capabilities, especially about distances. Further, success at these tasks requires the LLM to have been exposed to the relevant places and to successfully disambiguate their spatial interpretations, which is affected by the popularity of the place name. To address these problems, we develop a novel approach using spatial Retrieval Augmented Generation (RAG) to enable an LLM to use relevant distance information not explicitly learned during training when answering spatial questions that involve distance reasoning. Our method encodes a spatial graph of geodesic and road network distances between cities and towns, as well as a mapping to their alternative names, and compares a variety of techniques to provide the LLM with the appropriate spatial context from those knowledge stores. By incorporating relevant spatial information, our method significantly outperforms the baseline LLM on several tasks requiring distance-based spatial reasoning, including nearest-neighbor questions and distance analogy questions. We further show through experiments which graph storage and retrieval methods perform best, and which are brittle when applied to more challenging questions. Given the vast array of possible places an LLM could be asked to reason about, our work offers a first step towards providing a rudimentary distance-oriented spatial ‘world model’ to complement the vast linguistic knowledge held in LLMs.
The rapid adoption of location-aware devices, such as smartphones, smartwatches, in-vehicle navigation systems, and asset-tracking tags, has resulted in massive volumes of spatio-temporal trajectory data. This data drives a wide range of commercial, transportation, and government applications. However, benchmarking these applications at scale remains difficult: existing network-based trajectory generators are batch-oriented, lack scalability, and cannot produce long-running, continuous trajectory streams that reflect realistic traffic dynamics. To overcome these limitations, we present TraSGen, a distributed, scalable, and continuous trajectory stream generator designed specifically for large-scale benchmarking of trajectory-driven systems. TraSGen produces realistic network-constrained trajectories using different traffic flow models that incorporate real-time road conditions, nearby moving and stationary objects, and other contextual factors. It accepts a GeoJSON-based road network as input, where nodes and edges may contain rich attributes. These attributes, combined with dynamic traffic information, guide the prediction of each subsequent trajectory point. TraSGen is adaptive; it can trade off the volume and synchronization level of traffic-data updates for higher generation speed, enabling users to benchmark systems under diverse load and realism requirements. Built on a shared-nothing architecture, it scales horizontally to support large and continuous trajectory workloads. Extensive experimental evaluation demonstrates TraSGen’s scalability, efficiency, and significant advantages over state-of-the-art trajectory generators.
Effective tourist route generation can substantially enhance visitor experiences in urban areas by offering a diverse set of relevant route options while promoting the efficient use of urban infrastructure. However, the generation of tourist routes poses substantial challenges, primarily due to a complex set of often conflicting optimization objectives, and the scarcity of historical user trajectories for training data-driven models. In this article, we propose DIRECT, a geospatially informed novel deep reinforcement learning-based approach for tourist route generation. DIRECT dynamically generates a diverse set of high-quality route alternatives, while respecting user-specified spatio-temporal constraints and category preferences. Crucially, DIRECT does not require any historical user trajectories; instead, it leverages publicly available point-of-interest (POI) data and road network information. The results of our extensive evaluation demonstrate that DIRECT outperforms baseline methods in generating diverse tourist routes.
Understanding human mobility patterns in urban areas is crucial for a wide range of applications, including urban planning, commercial area analysis, tourism strategies, and disease control measures. In response to these diverse needs, several area embedding techniques have been actively researched, aiming at modeling mobility patterns for individual urban areas as vector representations within latent spaces. On the other hand, word embeddings are known to exhibit a property called additive compositionality, where arithmetic operations on embeddings correspond to meaningful semantic manipulations (e.g., “king” - “man” + “woman” ≈ “queen”). However, it remains unclear whether this property also holds in area embeddings, and if so, how it can be practically applied. Therefore, this study introduces and formulates the concept of additive compositionality from linguistics into area embeddings, defining it as a frequency-weighted average where the weights correspond to the frequency of occurrence for each area in the mobility dataset. Furthermore, we propose practical methods leveraging this property for flexible post-hoc transformations of area embeddings into different spatial shapes without retraining, and quantitative modeling of temporal mobility shifts as arithmetic operations between embeddings from different periods. Our results demonstrate that additive compositionality in area embeddings can effectively support operations on mobility patterns, spatial embedding transformations, and analysis of temporal mobility dynamics.
We address the problem of efficient detection of clusters of co-moving objects in a collection of molecular trajectories which, in addition to proximity , also satisfy semantic criteria in terms of chemical properties of interest. Specifically, we are interested in the motion of atoms from different molecules that at some point of the motion form Hydrogen Bonds (HB) and that HB persists over time. While a traditional continuous spatial proximity is an important criterion, the semantic aspect of which atoms from which molecules are co-moving, and whether they are also within certain spatial bounds is just as important. It may be tempting to use existing formalisms such as flocks, convoys, swarms, and so on—however, there are notable differences due to the semantics of the chemistry of HB formation: (1) There are additional constraints within clusters; and (2) From the perspective of the chemical interactions, it is permissible that an HB within a cluster is disrupted for a brief period (i.e., the persistency of the bond can be relaxed), for as long as it is re-established again soon. To enable the detection of such phenomena in datasets of atomic trajectories, we introduce the notion of Bond-Aware Relaxed Moving Clusters (BARMC) pattern—a novel type of spatio-temporal moving cluster pertaining to molecular dynamics. We provide an algorithm for the efficient detection of BARMC patterns along with experimental evaluation demonstrating its benefits over the Naïve approach based on traditional convoys.
Achieving carbon neutrality in personal mobility is dependent on technological improvements as well as considerable behavioural changes to reduce overall travel volumes and distances. However, implementing policies that aim to change behaviours also bear the risk of creating (more) inequalities. This paper proposes a combination of a new accessibility indicator with monitoring by mobile phone data analysis to accelerate and steer carbon reduction in transport. Minimum Mobility Standards (MMS) are a recently introduced accessibility indicator that defines the minimum carbon emissions required at a location to ensure access to essential daily activities. A comparison of these theoretical minimums with actual travel behaviour, as inferred from anonymised mobile phone data (MPD), and an estimation of transport modes and related CO2 emissions, reveals whether local policies should focus on steering behaviour change or improving accessibility to activities to enable change at first. The findings emphasise the potential for targeted policy interventions to further reduce emissions, particularly in medium-density zones where overshoot is most evident. Recommendations are provided for differentiated strategies based on urban form and population density.
We propose CLIFT ( C ross-City Lif estyle Pattern T ransfer for Human Mobility Prediction), a novel framework that enhances human mobility prediction by integrating general lifestyle patterns shared across cities with city-specific mobility patterns. Accurate human mobility prediction in urban environments is critical for transportation planning, marketing strategies, and disaster response. However, most existing deep learning approaches use only single-city data and exhibit significant performance degradation in small cities with limited training data. These limitations motivate methods that jointly leverage cross-city behavioral patterns and city-specific mobility characteristics. CLIFT addresses this challenge through dual complementary encoders: one captures general lifestyle patterns shared across cities, and the other captures city-specific mobility patterns; their outputs are integrated with a Transformer-based mobility predictor (LP-BERT). This architecture enables the model to jointly capture cross-city transferable behavioral patterns and city-specific mobility characteristics. We evaluated the effectiveness of CLIFT through experiments on the multi-city human mobility dataset LYMob-4Cities, comparing its performance with both single-city and multi-city deep learning-based methods. On average, CLIFT improved GEOBLEU and Top-1 accuracy by 11.1% and 10.6% over the single-city baseline, and by 5.0% and 7.9% over the multi-city baseline, respectively. Furthermore, CLIFT outperformed the top-ranked teams in the international competition, Human Mobility Prediction Challenge 2024 , demonstrating superior predictive performance under the same dataset and task setting.
The selection of key grid cells is critical for enhancing the interpretability and reliability of urban spatial-temporal flow prediction. Manual selection by domain experts can be inconsistent and time-consuming, while recent deep learning approaches that rely on feature visualization often produce unstable results due to sensitivity to input changes. In this work, we propose CRBF-Net, a convolutional radial basis function network that decouples feature representations from prediction outputs by introducing spatially localized basis kernels to prioritize selection stability. Our proposed method is evaluated on real-world telecom mobility data from Taipei, Taiwan. Experimental results demonstrate that CRBF-Net ensures high selection consistency and enhances the transferability of spatial insights across various downstream prediction models, including LSTM, ANN, and KNN. MAPE reductions were observed across all locations, with improvements of up to 8.24% compared to prior visualization-based selection strategies. Additionally, stability analysis via the Jaccard Index confirms that CRBF-Net maintains significantly higher spatial consistency across multiple independent trials. These results highlight the robustness and practical utility of CRBF-Net in selecting meaningful spatial regions, providing a stable and generalizable approach to key grid cells selection in urban flow prediction while contributing to the development of trustworthy and interpretable GeoAI systems.
Spatiotemporal machine learning models are increasingly central to geographic information science. Still, despite their growing impact, existing model documentation practices remain insufficient for capturing the spatial, temporal, and contextual dependencies that critically affect model validity, transferability, and responsible use. In this vision paper, we propose Spatiotemporal Model Cards (STMCs), a domain-agnostic extension of the Model Cards paradigm tailored to the unique characteristics of spatiotemporal models. The STMC framework integrates geographic and temporal metadata, autocorrelation-aware evaluation protocols, performance variability across space and time, transferability considerations, and ethical and sustainability aspects. Beyond model documentation, we outline how STMCs can serve as actionable interfaces within future-proof GeoAI systems, supporting model discovery, assessment, and reuse in both catalog-based and agentic (LLM-driven) settings. Using illustrative examples from the mobility domain, we demonstrate how STMCs help surface uncertainty, bias, and contextual limitations that are often obscured by aggregate performance metrics. We conclude by identifying key research challenges and directions for the GIScience community. STMCs provide a foundation for more transparent, reproducible, and responsible GeoAI practices.
Advances in artificial intelligence (AI) and multimodal sensing are driving a paradigm shift in the geospatial sciences, moving from task specific GeoAI models toward general purpose Geospatial Foundation Models (GeoFMs). While these models offer unprecedented opportunities for Earth monitoring, geographic knowledge discovery and addressing societal challenges such as natural disaster management, challenges remain regarding multimodal alignment, spatial reasoning, spatial distribution shifts, and generalizability. This article introduces the first section of a special issue dedicated to advancing the state-of-the-art in GeoFMs and their applications, featuring diverse research on neurosymbolic AI, heterogeneous graph learning, self-supervised learning, spatial retrieval-augmented generation, a genealogical review, and additive compositionality in urban representation learning. Together, these contributions offer new insights toward the next generation of geospatial intelligence.
Change detection in multi-temporal satellite data is critical for climate resilience assessment. Despite significant advances in remote sensing technologies, a fundamental research gap persists in effectively handling irregularly collected satellite images, where varying acquisition intervals and atmospheric conditions compromise detection accuracy. This study addresses this gap by developing a robust methodology for identifying changed and unchanged regions in multi-temporal satellite images. Our objective is to create a change detection framework that detects changed and unchanged regions in satellite images having temporal inconsistencies and inherent class imbalance. We propose a novel multiview-multitask learning in change detection (MMLCD) approach that integrates information from both frequency and spatial domains. The frequency domain analysis captures global structural changes while minimizing the impact of temporal variations, whereas the spatial domain processing through a CNN encoder-decoder identifies fine-grained localized changes. We enhance this framework with adversarial training and develop a custom loss function to address the class imbalance, the imbalance between hard and easy samples, and the instability in adversarial training. The proposed method is validated on benchmark datasets collected from Landsat-8 over nine diverse locations and the Multi-Temporal Urban Development SpaceNet (MUDS) dataser; the result indicates the superior performance of MMLCD over several state-of-the-art. We applied our model to Sentinel-2 satellite imagery of Nioghalvfjerdsfjorden Glacier in Greenland to assess its robustness and practical applicability in real-world scenarios where reliable annotations are unavailable, and to validate our approach’s effectiveness in monitoring critical environmental processes. The implications of this research extend beyond technical advancement. It includes practical applications in disaster response, and sustainable development planning, offering policymakers more reliable tools for environmental monitoring and informed decision-making in the face of accelerating global changes.
Traffic speed prediction is vital in fostering sustainable and efficient transportation systems, optimizing urban mobility, and improving road safety. An accurate prediction of traffic speed is challenging, especially for long-term predictions. One critical challenge in this context is the difficulty of capturing complex spatial and temporal interactions within road networks and making them accessible for deep learning models. Whereas recent prediction methods often rely on latent embedded representations of road networks, conventional latent spaces learn time-invariant representations of spatio-temporal snapshots and do not appropriately capture complex spatial and temporal dynamics. Furthermore, traffic data is often enriched by contextual information, such as features extracted from external sources, to enhance prediction accuracy. However, external contextual information sources for specific locations and times are typically limited. This article proposes a novel, neighborhood-based self-enrichment approach for traffic speed prediction to address these limitations. Our approach effectively identifies and explicitly models spatio-temporal correlations and dependencies in the traffic data. Then, our approach leverages these patterns to enrich the data, making such patterns explicitly accessible to the prediction models. We evaluate our method on real-world datasets and demonstrate that our method outperforms the baselines across all considered datasets on average by 4.10% in terms of root mean squared error for traffic speed prediction.
The rapid growth of urban mobility and logistics demands innovative solutions to address congestion, inefficiencies, and sustainability challenges. Unmanned Aerial Systems (UAS) offer a promising alternative for urban transportation, particularly in last-mile deliveries and emergency response. To accommodate the anticipated scale of UAS operations, potentially involving millions of daily flights, the UAS Traffic Management (UTM) ecosystem must support safe, efficient, and fair airspace access while minimizing disruption. UAS operations in dense urban areas must be supported by efficient and scalable path planning and deconflicting mechanisms which compute collision-free trajectories. A dynamic rerouting mechanism allows UAS trajectories to dynamically adapt to real-time deviations caused by environmental factors, communication errors, or dynamic airspace constraints. In this paper, we introduce a UAS Path Planning with Dynamic Rerouting framework that leverages a space-time graph model to ensure safe and efficient variable speed UAS navigation in urban environments. Our approach provides collision-free trajectories by construction and enables real-time path adjustments by dynamically rerouting UAS in case of slowdown events. By incorporating real-time request management and collision-free trajectories, our model enhances the feasibility of integrating UAS into multimodal urban transportation networks. We analyze the computational efficiency of our approach and demonstrate its scalability through simulation experiments using a real-world urban map in very dense scenarios with up to 14000 UAS vehicles. Our results indicate that dynamic rerouting with collision-free trajectory computation improves UAS route availability while maintaining airspace safety and operational efficiency.
The advent of large language models (LLMs) has enabled powerful applications across several domains such as science, healthcare, finance, and law. However, the spatial inference capabilities of LLMs are limited. Our goal is to facilitate more accurate LLM responses to spatial queries. To this end, we leverage inference-time retrieval augmented generation (RAG) to enrich LLM context using external data. We present SpaRAGraph, a framework that i) performs spatial-to-RDF data processing to capture spatial relations between nearby entities, ii) indexes relation-RDFs using a graph to facilitate semantic traversal, and iii) retrieves the relevant context to a question at inference time, contextualizing it with factual, spatial information enhancing the LLM’s accuracy. Additionally, we present a spatial reasoning benchmark that challenges LLMs on binary, multiclass and multilabel classification tasks on real-world, spatial entities. Overall, SpaRAGraph sets the ground for using spatial knowledge retrieval techniques to improve LLM effectiveness in spatial reasoning tasks.
Road inspection and cleaning are crucial to securing driving safety. Deploying a fleet of robots that run through a city can inspect and clean pavements without causing road closure. To achieve high coverage, one has to prevent robots from going through a road segment more than necessary. However, robots may need more than one visit to a particular road segment to inspect a defect. The uncertain success rate of defect inspection and the unknown maximum number of defects hinder the efficacy. Such uncertainty and constraints in objectives can also be seen in security patrolling, trip planning, and network maintenance. We target the problem of multi-agent scheduling with bounded objectives. The scheduling aims for maximum road network coverage while ensuring sufficient visits to particular road segments for defect identification of an uncertain subject, such as potholes and faded markings during road inspection or crimes and parking violations during security patrolling. We leverage an approximate bi-objective algorithm and propose a hierarchical circular route-planning algorithm. Our approach maximizes the road coverage among robots and decreases the search space when maximizing defect identification. We evaluate our approach on four real-world datasets. Our approach outperforms comparative methods by 11% in multi-objective optimization.
The rapid adoption of drones across various domains, alongside advancements in computer vision, has driven growing interest in vision-based airborne object detection from moving aerial platforms. However, this task remains challenging due to the small scale of objects, camouflage within cluttered backgrounds, and occlusions. To address these challenges, we introduce an end-to-end detection framework that integrates a Drone Receptive Field Block (DRFB) to extract multiscale and geometrically diverse features, specifically designed to enhance the detection of small and camouflaged airborne objects. To model motion patterns over time while preserving spatial structure, particularly for detecting camouflaged, cluttered and occluded objects with limited appearance cues, we incorporate a Convolutional Long Short-Term Memory (ConvLSTM) module, which effectively captures temporal dependencies across consecutive frames. Additionally, we introduce a SpatioTemporal Attention Block (STAB), inspired by Multi-Head Attention, to aggregate spatial and temporal context for improved semantic understanding. The detection head combines a Swin Transformer with a Cross Stage Partial (CSP) Bottleneck, offering lightweight yet powerful global context modeling for robust detection in complex aerial scenes. We evaluate our model on four publicly available airborne object detection datasets from moving drones, achieving significant improvements in accuracy while maintaining real time inference speed. Moreover, when integrated into various You Look Only Once (YOLO) architectures, our spatial feature extraction module (DRFB) consistently boosts performance, demonstrating its broad applicability and effectiveness. The code is available online here.
Traffic congestion in urban road networks leads to longer trip times and higher emissions, especially during peak periods. While the Shortest Path First (SPF) algorithm is optimal for a single vehicle in a static network, it performs poorly in dynamic, multi-vehicle settings, often worsening congestion by routing all vehicles along identical paths. We address dynamic vehicle routing through a multi-agent reinforcement learning (MARL) framework for coordinated, network-aware fleet navigation. We first propose Adaptive Navigation (AN), a decentralized MARL model where each intersection agent provides routing guidance based on (i) local traffic and (ii) neighborhood state modeled using Graph Attention Networks (GAT). To improve scalability in large networks, we further propose Hierarchical Hub-based Adaptive Navigation (HHAN), an extension of AN that assigns agents only to key intersections (hubs). Vehicles are routed hub-to-hub under agent control, while SPF handles micro-routing within each hub region. For hub coordination, HHAN adopts centralized training with decentralized execution (CTDE) under the Attentive Q-Mixing (A-QMIX) framework, which aggregates asynchronous vehicle decisions via attention. Hub agents use flow-aware state features that combine local congestion and predictive dynamics for proactive routing. Experiments on synthetic grids and real urban maps (Toronto, Manhattan) show that AN reduces average travel time versus SPF and learning baselines, maintaining 100
Scalable general-purpose representations of the built environment are crucial for geospatial artificial intelligence applications. This paper introduces S2Vec, a novel self-supervised framework for learning such geospatial embeddings. S2Vec uses the S2 Geometry library to partition large areas into discrete S2 cells, rasterizes built environment feature vectors within cells as images, and applies masked autoencoding on these rasterized images to encode the feature vectors. This approach yields task-agnostic embeddings that capture local feature characteristics and broader spatial relationships. We evaluate S2Vec on several large-scale geospatial prediction tasks, both random train/test splits (interpolation) and zero-shot geographic adaptation (extrapolation). Our experiments show S2Vec's competitive performance against several baselines on socioeconomic tasks, especially the geographic adaptation variant, with room for improvement on environmental tasks. We also explore combining S2Vec embeddings with image-based embeddings downstream, showing that such multimodal fusion can often improve performance. Our findings highlight how S2Vec can learn effective general-purpose geospatial representations of the built environment features it is provided, and how it can complement other data modalities in geospatial artificial intelligence.
Foundation models have garnered increasing attention for representation learning in remote sensing. Many such foundation models adopt approaches that have demonstrated success in computer vision with minimal domain-specific modification. However, the development and application of foundation models in this field are still burgeoning, as there are a variety of competing approaches for how to most effectively leverage remotely sensed data. This paper examines these approaches, along with their roots in the computer vision field. This is done to characterize potential advantages and pitfalls, while outlining future directions to further improve remote sensing-specific foundation models. We discuss the quality of the learned representations and methods to alleviate the need for massive compute resources. We first examine single-sensor remote foundation models to introduce concepts and provide context, and then place emphasis on incorporating the multi-sensor aspect of Earth observations into foundation models. In particular, we explore the extent to which existing approaches leverage multiple sensors in training foundation models in relation to multi-modal foundation models. Finally, we identify opportunities for further harnessing the vast amounts of unlabeled, seasonal, and multi-sensor remote sensing observations.
