Maria Despoina Siampou

Recent advances in large foundation models (FMs) have enabled learning general-purpose representations in natural language, vi- sion, and audio. Yet geospatial artificial intelligence (GeoAI) still lacks widely adopted foundation models that generalize across tasks that require joint reasoning over geospatial objects and human mo- bility. Such tasks are crucial as mobility, along with satellite imagery, street view, and text, is a core modality for understanding the phys- ical world. We argue that a key bottleneck is the absence of unified, general-purpose, and transferable representations for geospatially embedded objects (GEOs). Such objects include points, polylines, and polygons in geographic space, enriched with semantic context and critical for geospatial reasoning. Much current GeoAI research compares GEOs to tokens in language models, where patterns of human movement and spatiotemporal interactions yield contextual meaning similar to patterns of words in text. However, modeling GEOs introduces challenges fundamentally different from language, including spatial continuity, variable scale and resolution, temporal dynamics, and data sparsity. Moreover, privacy constraints and global variation in mobility further complicates modeling and gen- eralization. This paper formalizes these challenges, identifies key representational gaps, and outlines research directions for building foundation models that learn behavior-informed, transferable rep- resentations of GEOs from large-scale human mobility data, as well as static contextual information such as points of interest, object shapes and spatio-temporal semantics.

Encoding geospatial data is crucial for enabling machine learning (ML) models to perform tasks that require spatial reasoning, such as identifying the topological relationships between two different geospatial objects. However, existing encoding methods are limited as they are typically customized to handle only specific types of spatial data, which impedes their applicability across different downstream tasks where multiple data types coexist. To address this, we introduce Poly2Vec, an encoding framework that unifies the modeling of different geospatial objects, including 2D points, polylines, and polygons, irrespective of the downstream task. We leverage the power of the 2D Fourier transform to encode useful spatial properties, such as shape and location, from geospatial objects into fixed-length vectors. These vectors are then inputted into neural network models for spatial reasoning tasks. This unified approach eliminates the need to develop and train separate models for each distinct spatial type. We evaluate Poly2Vec on both synthetic and real datasets of mixed geometry types and verify its consistent performance across several downstream spatial reasoning tasks.

Accurately modeling and analyzing time series data is crucial for downstream applications across various fields, including healthcare, finance, astronomy, and epidemiology. However, real-world time series often exhibit irregularities such as misaligned timestamps, missing entries, and variable sampling rates, complicating their analysis. Existing approaches often rely on imputation, which can introduce biases. A few approaches that directly model irregularity tend to focus exclusively on either capturing intra-series patterns or inter-series relationships, missing the benefits of integrating both. To this end, we present WaveGNN, a novel framework designed to directly (i.e., no imputation) embed irregularly sampled multivariate time series data for accurate predictions. WaveGNN utilizes a Transformer-based encoder to capture intra-series patterns by directly encoding the temporal dynamics of each time series. To capture inter-series relationships, WaveGNN uses a dynamic graph neural network model, where each node represents a sensor, and the edges capture the long- and short-term relationships between them. Our experimental results on real-world healthcare datasets demonstrate that WaveGNN consistently outperforms existing state-of-the-art methods, with an average relative improvement of 14.7% in F1-score when compared to the second-best baseline in cases with extreme sparsity. Our ablation studies reveal that both intra-series and inter-series modeling significantly contribute to this notable improvement.

The abundance of vehicle trajectory data offers a new opportunity to compute driving routes between origins and destinations. Current graph-based routing pipelines, while effective, involve substantial costs in constructing, maintaining, and updating road network graphs to reflect real-time conditions. In this study, we propose a new trajectory-based routing paradigm that bypasses current workflows by directly utilizing raw trajectory data to compute efficient routes. Our method, named TrajRoute, uniquely "follows" historical trajectories from a source to a destination, constructing paths that reflect actual driver behavior and implicit preferences. To supplement areas with sparse trajectory data, the road network is also incorporated into TrajRoute’s index, and tunable parameters are introduced to control the balance between road segments and trajectories, ensuring a unified and adaptable routing approach. We experimentally verify our approach by comparing it to an existing online routing service. Our results demonstrate that as the number of trajectories covering the road network increases, TrajRoute produces increasingly accurate travel time and route length estimates while gradually eliminating the need to downgrade to the road network. This highlights the potential of simpler, data-driven pipelines for routing, offering lower-maintenance alternatives to conventional systems.