Modeling Urban Intelligence

Abstract

Urban environments operate as layered systems in which human activities, built structures, and atmospheric processes interact continuously and nonlinearly. However, existing urban intelligence methodologies remain fragmented, often depending on specialized pipelines with limited sensing modalities, region-specific infrastructure, or short-range temporal reasoning. Such constraints restrict scalability, delay decision-making, and reduce the ability of models to generalize across heterogeneous urban contexts. In response, this dissertation develops the key methodological components required to work toward one—advancing large-scale environmental perception, scalable weather forecasting, and long-horizon generative prediction as complementary capabilities for future context-aware urban intelligence.

The first component of this work addresses the challenge of understanding social risks and human–environment interaction using incomplete and unevenly distributed urban sensing. We develop an approach that extracts meaningful representations from street-level imagery and interprets how physical space contributes to behavioral patterns within cities. Rather than relying solely on explicit records of crime or human annotation, the model learns how visual and contextual cues relate to observable social outcomes and how these relationships vary across different cultural and geographical settings. The resulting framework generalizes beyond its training context and remains functional even in regions where structured safety data are scarce, enabling practical use in applications such as safety-aware routing, environmental assessment, and data-informed urban decision-making.

The second component of this work focuses on building a weather intelligence capability that can operate effectively without dependence on specialized infrastructure. Rather than relying on traditional sensing systems or region-specific data pipelines, this approach learns weather-related patterns from broadly available observations and translates them into actionable forecasts. By emphasizing adaptability across regions with differing data quality, resource availability, and environmental conditions, the framework supports timely prediction of impactful events such as heavy precipitation and tropical cyclone movement. This enables forecasting systems to be deployed in contexts where conventional meteorological resources are limited, ultimately enhancing disaster preparedness and resilience in vulnerable regions.

The final component explores how long-term temporal structure can be modeled to support stable reasoning over extended sequences relevant to both environmental and urban processes. This part of the work focuses on developing a forecasting approach that can learn from gradual change as well as rare disruptive events, while maintaining interpretability and robustness over time. Rather than prioritizing short-term prediction alone, the framework emphasizes continuity, memory, and representational stability, allowing it to anticipate future states over long horizons. This capability enables more reliable modeling of evolving systems and supports applications that require sustained temporal awareness beyond the limits of conventional forecasting methods.

Together, these contributions illustrate that multi-modal sensing, adaptive learning, and scalable temporal modeling can be treated as compatible and complementary components within the broader field of urban intelligence. Rather than functioning as isolated technical solutions, the methods developed in this dissertation demonstrate how social perception, environmental forecasting, and long-horizon reasoning can be aligned conceptually and methodologically. In doing so, this work establishes practical and transferable building blocks that can be integrated into more comprehensive systems aimed at interpreting, predicting, and responding to the evolving dynamics of urban environments.