GeoAI has evolved from an interdisciplinary niche into a core paradigm for spatial environmental modeling. By embedding spatial dependence, context, and scale directly into learning architectures, GeoAI extends classical spatial analysis beyond descriptive statistics toward predictive environmental intelligence (Gao, 2023; Huang, 2025). While traditional geostatistical methods such as Moran’s I and LISA established the foundations of spatial reasoning (Moran, 1950; Anselin, 1995), their assumptions of stationarity and linearity limit applicability in heterogeneous urban landscapes.

Modern GeoAI architectures learn nonlinear, multiscale interactions among climatic, demographic, and built-environment variables (multi-modality), enabling high-resolution mapping of land-surface temperature (LST) and microclimate variability (Fu et al., 2025; Mukarram et al., 2025). However, most existing applications remain task-specific, episodic, and city-bound, relying on customized models trained for individual study areas.

Recent work in Geospatial Explainable AI (XAI) emphasizes the need for interpretable spatial reasoning, particularly in environmental and public-health contexts where trust and transparency are critical (Roussel and Böhm, 2023). At the same time, hybrid approaches integrating physical constraints—such as surface energy balance—have emerged to improve generalization across climatic regimes (Yoo and Weng, 2024).

Despite these advances, most GeoAI applications remain task-specific and episodic, relying on locally trained models that are difficult to generalize across cities or seasons. Moreover, spatial reasoning is often learned implicitly rather than embedded through physically informed or transferable representations, limiting the operational scalability of GeoAI for continuous urban climate monitoring.

Machine-learning approaches have substantially improved the detection and mapping of urban heat islands by capturing nonlinear interactions among surface materials, vegetation, and urban morphology (Yoo and Weng, 2024; Li et al., 2024). Deep learning and ensemble methods consistently outperform classical regression techniques for intra-urban temperature estimation, particularly at sub-kilometer scales.

Nevertheless, most urban heat models are trained on limited temporal windows and localized datasets, constraining transferability across cities and climatic regimes (Koumetio Tekouabou et al., 2022). While recent studies have explored data fusion across satellite platforms and socioeconomic indicators (Johnson et al., 2009; Johnson et al., 2012; Li et al., 2024), these pipelines remain largely episodic and decoupled from atmospheric dynamics.

As a result, existing GeoAI-based heat-mapping approaches struggle to provide temporally continuous, physically consistent representations of urban thermal behavior, particularly when applied beyond the spatial or temporal domain of the training data.

AI-based health-risk models increasingly rely on high-resolution environmental exposure estimates to characterize heat-related morbidity and mortality (Ho et al., 2025). GeoAI enhances this linkage by introducing spatial continuity and contextual reasoning into exposure assessment, addressing limitations of station-based temperature observations that fail to capture neighborhood-scale variability.

However, most health-focused applications treat exposure surfaces as fixed inputs rather than as modeled products whose spatial resolution, temporal continuity, and physical realism directly condition downstream inference. Consequently, uncertainty in exposure estimation propagates into health-risk models, often without explicit consideration of how those exposure fields were generated.

This disconnect highlights the need for exposure-centric GeoAI pipelines that prioritize the generation of physically consistent, high-resolution urban heat fields suitable for integration into health-risk analytics.

Across the GeoAI literature on urban heat and heat-risk mapping, three unresolved challenges recur: limited temporal continuity in surface temperature observations, constrained transferability of city-specific machine-learning models, and weak integration between atmospheric and surface representations. While foundation models and multimodal embeddings offer a promising pathway toward addressing these limitations, few studies have demonstrated how such architectures can be operationalized in practice.

The present study helps to address this gap by implementing a multimodal GeoAI pipeline that fuses Prithvi-WxC atmospheric embeddings with Prithvi-EO surface representations and Granite-LST translation to generate hourly, super-resolved urban heat fields. The following section presents a case-study implementation in Indianapolis, Indiana, demonstrating the feasibility of foundation-model-based urban heat mapping as a precursor to downstream health-risk assessment.

Indianapolis, Indiana, represents an archetypal mid-sized U.S. city facing an intensifying threat from extreme heat under climate change. The region experiences increasing frequency and duration of heatwaves, with urban heat islands producing intra-city temperature differences exceeding 5 °C between dense commercial cores and vegetated suburbs (Liang and Weng, 2008). These spatial and temporal variations in surface temperature strongly influence local exposure patterns and contribute to health disparities observed in emergency-department visits and heat-related hospitalizations (Beeharry et al., 2025; Ho et al., 2025).

Despite abundant meteorological and remote-sensing data, the ability to monitor heat exposure continuously at high spatial and temporal resolution remains limited. Satellite sensors such as MODIS and Landsat offer complementary strengths (temporal coverage versus spatial fidelity), but their combined cadence yields at most a few clear-sky observations per week. This gap constrains epidemiological studies and real-time decision support.

To help address this limitation, the PrithviWxC–PrithviEO–GraniteLST GeoAI pipeline was implemented, which fuses foundation-model representations of atmospheric and surface conditions to generate hourly, super-resolved LST estimates for the Indianapolis region.

The pipeline links three large-scale GeoAI models (Figure 1):

Prithvi-WxC (Schmude et al., 2024) (Weather x Climate Foundation Model): Provides hourly (variable temporal resolution) latent representations of atmospheric state variables, temperature, humidity, wind, and radiation, derived from reanalysis and numerical weather prediction fields.

Workflow Summary

Step 1 – Data Ingestion: ERA5 reanalysis and short-term forecast products feed Prithvi-WxC; HLS imagery and static layers feed Prithvi-EO.

To integrate heterogeneous environmental information while preserving spatial structure, a latent space fusion head was employed downstream of modality-specific encoders. Rather than performing early fusion at the pixel or input level, fusion was conducted on aligned latent representations, enabling context-aware weighting of Earth observation and meteorological information.

Multispectral surface reflectance inputs derived from Harmonized Landsat–Sentinel (HLS) imagery were first mapped into a shared latent feature space using a pretrained Granite land-surface temperature (LST) foundation model. This encoder produces spatially explicit latent feature maps that retain the original grid topology and georeferencing of the input tile. Auxiliary meteorological context, represented by the spatial mean of ERA5-Land 2-m air temperature at acquisition time, was incorporated as a scalar side channel and projected into the latent space via a learnable linear transformation. This scalar embedding was subsequently broadcast across the spatial dimensions of the latent field, allowing large-scale atmospheric context to modulate local predictions without imposing artificial spatial gradients

The fusion head aligns all modality-specific latent representations to a common channel dimensionality using 1 × 1 convolutions, after which they are combined through a learned gating mechanism. This module produces spatially varying weights for each modality, enabling adaptive emphasis of Earth observation features or meteorological context depending on local conditions. The weighted latent fields are then aggregated and passed through a shallow convolutional refinement block to produce the final fused representation used for prediction.

This design yields several advantages. Fusion occurs in a representation space already structured by the Granite foundation model, reducing the need for extensive retraining while leveraging pretrained physical and spectral relationships. Scalar climate information influences predictions in a globally consistent but locally expressive manner, avoiding the need for full meteorological raster inputs when such data are unavailable or uncertain. Finally, the spatially adaptive gating mechanism provides a principled alternative to static concatenation or fixed-weight fusion, allowing the relative influence of each data source to vary across space.

The pipeline operates through a scripted PowerShell/Python workflow controlling tiled inference across the Indianapolis AOI. Model configurations and checkpoints for Granite-LST are loaded dynamically; each run iterates through 24 h per target day, generating hourly outputs.

Data are streamed in chunked Zarr format, enabling out-of-core processing and scalability to regional domains. Experiments run on a multi-GPU Linux cluster (4 × A100 80 GB) using PyTorch 2.4 and mixed-precision inference.

To ensure reproducibility, each module logs configuration parameters, tile indices, and validation metrics to an MLflow instance. Common failure modes, e.g., tensor-shape mismatches, zero-filled outputs from mis-wired inputs, or overlap errors in tiled inference, were systematically debugged and resolved during prototype development.

To demonstrate the operational feasibility and physical consistency of the proposed GeoAI framework, this section presents a focused case study centered on a clear-sky summer day in Indianapolis. Accordingly, the analysis is intentionally scoped as a proof-of-concept demonstration rather than as a comprehensive assessment of cross-seasonal or cross-city generalization. The fused PrithviWxC–PrithviEO–GraniteLST pipeline generated continuous hourly LST fields for 1 July 2024 under clear-sky conditions. Figure 2 illustrates the modeled diurnal temperature evolution averaged across three representative surface classes—urban core, suburban, and vegetated zones—together with the temperature differential (ΔT) between urban and vegetated areas.

Early-morning minima near 07 h EDT (≈31 °C urban; 21 °C vegetated) precede rapid heating after 08 h, with peak LSTs approaching 41 °C in dense impervious zones by mid-afternoon. The modeled ΔT (urban–vegetated) widens sharply during morning warming and stabilizes near 13 °C between 14 and 17 h, reflecting realistic intra-urban heat-island behavior consistent with in-situ, Landsat, and ASTER observations.

Figure 3 depicts the corresponding spatial fields for twelve representative hours (07–18 h EDT). The transition from green to red hues highlights the pronounced diurnal amplification of thermal gradients, especially along industrial corridors and downtown Indianapolis. Between 11 and 17 h EDT the model captures both macro-scale atmospheric forcing and fine-scale heterogeneity at ∼30 m resolution. Mean domain temperature rises from 26.6 °C at 07 h EDT to 34.8 °C by 17 h EDT, then stabilizes, consistent with the latent heat capacity of impervious surfaces.

Spatially, the hottest sectors coincide with the I-465 bypass loop and urban infill zones lacking tree canopy, while cooler microclimates appear along the White River and parkland buffers. The model reproduces key physical signatures of diurnal heat evolution, early-afternoon plateau, delayed urban cooling, and spatially coherent thermal inertia.

This case study demonstrates the feasibility of integrating atmospheric and land-surface foundation models within a unified GeoAI pipeline to generate continuous, hourly land-surface temperature fields at neighborhood scale. By coupling Prithvi-WxC–derived atmospheric context with Prithvi-EO surface representations and a downstream Granite-LST predictor, the pipeline bridges a persistent observational gap between sparse satellite overpasses and the temporal resolution required for exposure-relevant heat analysis. The results show that multimodal latent integration can reproduce realistic diurnal thermal dynamics and fine-scale spatial heterogeneity without reliance on dense in-situ sensor networks or continuous clear-sky imagery. Importantly, the contribution of this work lies in methodological integration and operational feasibility rather than in the direct estimation of health outcomes, establishing a foundation for subsequent coupling with epidemiological and public-health models. Building on this methodological foundation, the ultimate purpose of this pipeline is to provide proof of concept in the development of dynamic exposure fields for heat-health modeling. Hourly LST maps can be linked with:

Electronic health-record (EHR) data to quantify exposure–response relationships;

Future work will integrate these exposure fields into spatial epidemiological models of heat-related hospitalizations (Toure et al., 2025) and mortality differentials (Beeharry et al., 2025). This coupling will enable near-real-time to real-time, neighborhood-scale estimation of health risk and inform equitable adaptation strategies.

While the multimodal GeoAI pipeline demonstrates strong performance and physical consistency, several research directions warrant continued development and validation. Although satellite imagery remains an important input, the fused foundation-model architecture overcomes the traditional clear-sky constraint of remote sensing. By conditioning on recent clear-sky observations and assimilating atmospheric states from Prithvi-WxC, the system can synthetically generate physically consistent LST fields for cloudy or sensor-gap periods. This capability represents a shift from observational interpolation to generative environmental modeling, enabling continuous thermal field estimation under data gaps.

Future research should prioritize cross-regional validation to assess the generalizability of synthetic outputs under differing climatic and urban contexts. Further coupling with federated health-risk models will allow privacy-preserving integration of clinical, environmental, and socio-economic data streams. Integrating additional hydrological, energy, and behavioral indicators could extend the framework toward compound-risk modeling (e.g., concurrent heat, humidity, and air-quality stressors). Finally, although beyond the scope of this study, embedding formal uncertainty quantification and explainable AI components will strengthen scientific transparency and user trust, particularly in health and policy settings.

Beyond scientific validation, realizing the full potential of this system requires translating the GeoAI pipeline into operational climate-health infrastructure. The PrithviWxC–PrithviEO–GraniteLST architecture transforms disparate meteorological and Earth-observation data into continuous, high-resolution thermal intelligence that supports both research and public-health action.

Scaling beyond Indianapolis involves several key implementation steps:

Extending tiled inference to regional and national domains through distributed computing;

Collectively, these steps advance the concept of AI-enabled urban climate twins, digital infrastructures capable of diagnosing, forecasting, and mitigating extreme-heat risk in real time. Such systems illustrate how foundation-model fusion can transform GeoAI from a research instrument into a practical, decision-support framework for climate adaptation and health resilience.