Physics-Guided Inductive Spatiotemporal Kriging for PM2.5 with Satellite Gradient Constraints

The gist The Spatiotemporal Physics-Guided Inference Network (SPIN) is a novel framework designed for inductive spatiotemporal kriging that synergistically integrates domain knowledge into deep learning by explicitly modeling physical advection and diffusion processes via parallel graph kernels, allowing it to learn structural pollution patterns from satellite data while remaining robust to data voids

Introduction and Problem Statement

Fine particulate matter (PM2.5) remains a formidable threat to global public health and ecological stability. While traditional data-driven methods attempt to bridge this gap using satellite Aerosol Optical Depth (AOD), they often suffer from severe, non-random data missingness due to cloud cover or nighttime and inversion biases. The research frontier must expand from fixed-point forecasting to inductive spatiotemporal kriging—the capability to reconstruct continuous, high-resolution pollution fields for arbitrary unobserved locations based on sparse point measurements. Current data-driven deep learning architectures often lack the inductive capability to generalize outside the training graph, relying on fixed graph structures defined during training.

SPIN Framework and Architecture

The SPIN framework distinguishes itself by embedding physical laws into the deep learning architecture and treating satellite data as a constraint rather than an input. The architecture consists of three hierarchical stages: temporal feature encoding, physics-guided propagation, and constrained optimization.

The model first employs a Temporal Convolutional Network (TCN) to encode local temporal dynamics for each node, mapping the input timeseries of meteorology and emissions to a high-dimensional latent representation H(0)i. It then propagates spatial information using two explicit kernels:

1. Geospatial Graph (GS): This graph captures the fundamental principle of spatial autocorrelation, constructing an adjacency matrix based on pairwise geographic distances using a Gaussian kernel.

2. Diffusion Graph (GD): Based on GS, this models the physical process of diffusion by defining a diffusion operator A˜D corresponding to the Laplacian operator on the graph.

3. Advection Graph (GA): This kernel explicitly models anisotropic transport driven by wind fields, where an edge exists only if node j is upwind, and its weight is determined by the projection of the wind vector onto the node-pair vector.

The propagation layer updates node representation H(l) by aggregating messages from both operators: H(l) = σ A˜DH(l−1) + AAH(l−1) W(l), where W(l) is the learnable weight matrix.

Paradigm-Shifting Training Strategy

A critical component of SPIN is the paradigm-shifting training strategy, which involves a composite loss function L = Linfer + λ1Linit + λ2LAOD.

The three components of the loss function are:

(a) Inference Loss (Linfer):

This is the primary objective, minimizing the L1 error on the masked target nodes (Vtarget) to simulate inference by combining propagated neighbor information with local features.

(b) Initialization Loss (Linit):

This loss forces the TCN to generate a reasonable initial proposal (Xˆinit) based solely on local meteorology and emissions before neighbor information is aggregated, providing a physics-based baseline estimate.

(c) Masked AOD Spatial Gradient Loss (LAOD):

This term introduces spatial structure constraints by aligning the spatial gradient of the final prediction with that of the satellite AOD field, ensuring robustness to missing data. The loss is defined as ∇ij (Xˆ) − ∇ij (XAOD) · Mt ij (11).

Results and Performance

The validation in the highly polluted Beijing-Tianjin-Hebei and Surrounding Areas (BTHSA) demonstrates that SPIN achieves a new state-of-the-art with a Mean Absolute Error (MAE) of 9.52 µg/m3.

The model shows superior performance in Station Inference, achieving an overall yearly MAE of 9.52 µg/m3, representing a substantial 25.2% improvement over the strongest baseline, IGNNK.

The framework demonstrates robustness under extreme data scarcity; for instance, in spatially isolated regions with minimal upstream sensors, the Diffusion Kernel acts as a conservative low-pass filter in data vacuums, ensuring inference remains physically plausible even when information is extremely scarce.

For Grid Inference across varying satellite data conditions, SPIN successfully reproduces complex spatial patterns when AOD and ground truth align, and it robustly falls back to its internal physics-informed kernels when satellite data is completely unavailable.

Discussion and Implications

The superior performance of SPIN validates the PhysicsEmbedded design philosophy by explicitly disentangling pollutant transport into isotropic diffusion and anisotropic advection, which captures the dominant wind-driven transmission corridors. The strategy of using satellite data as a loss constraint rather than an input feature effectively decouples the model’s inference capability from the availability of real-time satellite data. This enables a physics fallback mechanism, ensuring continuous operation when satellite eyes are blinded by clouds or darkness. The inductive inference capability of SPIN allows it to function as a cost-efficient virtual sensor, providing reliable air quality estimates for unmonitored suburbs and rural areas.

Conclusion

To address the practical challenge of spatiotemporal inference in unmonitored urban areas, this study proposes SPIN, a physics-guided inductive graph neural network. The framework successfully reconstructs high-resolution PM2.5 fields from sparse ground observations by innovating in three dimensions: Architecture, Strategy, and Data Fusion. This novel approach effectively solves the pervasive missing data problem caused by cloud cover, achieving all-weather inference capability. The work provides a scalable, low-cost, and robust solution for next-generation environmental sensing.

Acknowledgments

We extend our gratitude to the Ministry of Ecology and Environment of the People’s Republic of China, the Japan Meteorological Agency (JMA), and the European Centre for Medium-Range Weather Forecasts (ECMWF) for providing the essential data supporting this study. The work was supported by the National Natural Science Foundation of China (Grant Nos. 42450183, 12275020, 12135003, 12205025, and 42461144209) and the Ministry of Science and Technology of China (Grant No. 2023YFE0109000). The paper is available on arXiv under the identifier arXiv:2511.16013v1 [cs.LG] 20 Nov 2025

--- Page 7 ---

The model first encodes local temporal dynamics via a Temporal Convolutional Network (TCN) to encode the local temporal dynamics for each node. Compared to RNNs, TCNs employ dilated causal convolutions to capture long-range temporal dependencies more efficiently with parallel computing. This step provides a physics-informed initialization for all nodes, offering a robust starting point even for unmonitored grid cells where no historical pollution data is available.

--- Page 9 ---

We explicitly model the spatiotemporal dependencies using a set of three distinct graphs, each encoding a specific physical or statistical relationship.

The Geospatial Graph (GS) captures the fundamental principle of spatial autocorrelation (Tobler’s First Law of Geography).

The Diffusion Graph (GD) models the physical process of diffusion, where pollutants spread from high to low concentration areas, corresponding to the Laplacian operator on the graph <ref:2511.

1. textbfPhysics-Embedded Architecture for Physical Plausibility: The SPIN Framework Enables Modeling of Atmospheric Dynamics by explicitly modeling physical advection and diffusion processes via parallel graph kernels. This allows the model to learn spatiotemporal dependencies consistent with physical laws, specifically disentangling isotropic diffusion (local spreading) and anisotropic advection (directional transport).

2. textbf Masked AOD Spatial Gradient Loss for Robust Data Handling: The system utilizes a novel training strategy where it uses the spatial gradients of available AOD data to constrain the model’s output structure during training, which allows it to robustly leverage the structural patterns in satellite imagery while elegantly circumventing the pervasive issue of missing data.

3. textbf Inductive Spatiotemporal Kriging for Blind Spot Filling: The architecture is designed for inductive inference, enabling it to "accurately predict PM2.5 concentrations at any arbitrary unobserved location (Station Inference) and generate complete maps (Grid Inference), effectively extending prediction capabilities from discrete sites to the global space."

4. textbf Dynamic Node Masking Strategy for Generalization: The training incorporates a dynamic masking strategy where the model is forced to reconstruct the values of these masked nodes by combining their local meteorological and emission features with pollution context propagated from the remaining observed nodes, which prevents the model from overfitting to specific station IDs.

5. textbf All-Weather, Low-Cost Environmental Management: The resulting system provides a robust, low-cost, and all-weather solution for fine-grained environmental management by falling back on physics-driven extrapolation based on meteorology and ground observations when satellite data is unavailable (Physics Fallback).