��������

As wildfire seasons intensify and air pollution continues to threaten public health, geographers are turning to new generations of artificial intelligence models to understand how environmental hazards unfold across space. Geography Professor , PhD student , and collaborator of National Jewish Health are developing next-generation models that combine satellite data, atmospheric information, and AI-driven “place signatures” to better estimate air pollution across the United States.

Their latest publication accepted in focuses on PM₂.₅, a harmful form of air pollution linked to asthma, cardiovascular disease, and premature mortality. Traditional approaches rely on networks of ground-based monitors and satellite-derived aerosol data, but both leave important gaps. Many communities, especially in rural regions or areas affected by sudden wildfire smoke, lack reliable monitoring. Pollution also varies dramatically from one neighborhood to the next. This creates both scientific and equity challenges.

To address these gaps, the team built a that synthesizes 21 days of satellite observations, meteorological variables, wildfire smoke information, and other environmental data to estimate daily PM₂.₅ at high spatial resolution. The model is designed to follow how pollution evolves over time, capturing the dynamics of major smoke events and seasonal changes.��But their latest innovation adds something novel to the discipline: geospatial foundation models, including “location encoders” such as , incorporated in a practical way for dynamic air pollution estimation. These models learn from millions of ground-level photographs—urban streetscapes, forests, industrial landscapes, suburban neighborhoods—to produce rich, 512-dimensional embeddings that describe the visual and contextual character of places. When incorporated into the air-quality system, these learned representations provide information about land use, vegetation, density, and built environments that traditional datasets often miss.

“Location encoders give our models a deeper understanding of what a place is like,” says Karimzadeh. “They capture signals that satellites alone can’t see—traffic corridors, industrial zones, tree cover—and that helps us estimate pollution more accurately, especially in places with few monitors.”

The impact is clear in case studies like the 2021 Dixie Fire, when thick smoke blanketed large portions of the western U.S. Models enhanced with location embeddings captured not only the concentration of PM₂.₅ but also the full spatial extent of the smoke plume with greater precision and coherence than satellite-only approaches.

For Wang, who leads much of the model development, has been pursuing this goal of building models that generalize well and provide reliable information even where monitoring is sparse.

In the future, the team aims to incorporate additional sensors and imagery into their models, and explore seasonal and long-term place representations. Their research reflects a broader paradigm in geography and environmental science: using AI not to replace traditional observation methods, but to complement and strengthen them.

��

��

Figure from the published paper: Estimated PM2.5 during the 2021 Dixie Fire (Northern California) produced by the baseline model (without geographic features) and the GeoCLIP-enhanced model. Each row corresponds to a different day during peak wildfire activity. Columns (a) and (b) show baseline results, while columns (c) and (d) present GeoCLIP-enhanced estimates at both CONUS and regional scales. The GeoCLIP model yields more intense and spatially coherent smoke plumes and additionally identifies elevated PM2.5 levels over northern Minnesota on July 21, reflecting long-range smoke transport from simultaneous western U.S. and Canadian wildfires.