Foundation AI Models for Aerosol Optical Depth Estimation from PACE Satellite Data
| Authors | Zahid Hassan Tushar & Sanjay Purushotham |
| Year | 2026 |
| Field | Computer Vision |
| arXiv | 2605.00678 |
| Download | |
| Categories | cs.CV |
Abstract
Aerosol Optical Depth (AOD) retrieval is essential for Earth observation, supporting applications from air quality monitoring to climate studies. Conventional physics-based AOD retrieval methods formulate the problem as a pixel-wise inversion, relying on radiative transfer modeling, memory-intensive look-up tables, and auxiliary meteorological data. While recent data-driven approaches have shown promise, many fail to exploit the spatial-spectral coherence of hyperspectral imagery, leading to spatially inconsistent and noise-sensitive retrievals. We present the first study exploring Foundation AI models for AOD retrieval and propose ViTCG, a Vision Transformer with Channel-wise Grouping-based spatial regression framework that reduces retrieval bias and error. ViTCG uses hyperspectral top-of-atmosphere radiance as input and jointly models spatial context and spectral information. Validation with PACE radiance observations demonstrates a 62% reduction in mean squared error compared to state-of-the-art foundation models, including Prithvi, and produces spatially coherent AOD fields.
Engineering Breakdown
Plain English
This paper tackles Aerosol Optical Depth (AOD) retrieval from satellite imagery—a critical task for air quality monitoring and climate science—by proposing ViTCG, a Vision Transformer-based framework that exploits spatial-spectral relationships in hyperspectral data. Traditional physics-based methods rely on expensive radiative transfer models and lookup tables, while recent data-driven approaches ignore the spatial coherence patterns in images, producing noisy and inconsistent results. The authors demonstrate that foundation models (Vision Transformers) can significantly reduce retrieval bias and error when properly adapted with channel-wise grouping mechanisms to preserve spectral information across spatial regions. This represents the first systematic application of foundation AI models to AOD retrieval, showing that modern deep learning can outperform conventional satellite inversion methods.
Core Technical Contribution
The core novelty is ViTCG (Vision Transformer with Channel-wise Grouping), which adapts foundation Vision Transformer models for pixel-wise regression tasks that require both spatial consistency and spectral sensitivity. Unlike standard Vision Transformers that treat all channels uniformly, ViTCG groups channels to preserve hyperspectral structure while learning spatial dependencies across the image. This is the first work to systematically explore foundation AI models for AOD retrieval, moving away from physics-based lookup tables and memory-intensive radiative transfer models toward learned representations. The key insight is that spatial-spectral coherence—the fact that neighboring pixels with similar spectral signatures should have similar AOD values—can be implicitly learned by transformers better than hand-engineered physics models.
How It Works
ViTCG takes hyperspectral top-of-atmosphere (TOA) imagery as input, typically with dozens to hundreds of spectral bands across visible and near-infrared wavelengths. The Vision Transformer backbone processes the image patches, learning spatial attention patterns that capture which regions correlate with higher or lower aerosol optical depth. The Channel-wise Grouping mechanism groups the hyperspectral bands into semantically meaningful clusters (e.g., visible bands separate from SWIR bands) rather than treating all channels independently, preserving spectral coherence during feature learning. The spatial regression head then predicts a single AOD value per pixel, leveraging the learned spatial dependencies to reduce noise compared to independent pixel-wise predictions. The transformer's multi-head self-attention allows the model to learn which spectral combinations are most informative for AOD—automatically discovering that certain band ratios or spectral features correlate strongly with aerosol loading, similar to handcrafted spectral indices but in a learned, data-driven manner.
Production Impact
In operational Earth observation systems, replacing physics-based AOD retrieval with ViTCG would eliminate the need to maintain and update lookup tables (which require expensive radiative transfer simulations), reduce memory footprint on satellite processing systems, and improve consistency of AOD products across different sensor platforms. The spatial regression approach produces smoother, more geophysically plausible AOD maps compared to noisy pixel-wise inversion, directly improving downstream air quality forecasting and climate analysis. However, this approach requires substantial labeled training data (satellite images paired with ground-truth AOD from AERONET stations or similar)—a non-trivial constraint since high-quality aerosol validation data is sparse globally. Inference latency and compute requirements depend on image resolution and model size; processing full satellite swaths would require GPU acceleration, and the model must be retrained or fine-tuned for new sensor types (Sentinel-5P, TROPOMI, etc.), adding engineering overhead compared to universal physics-based models. Integration into existing operational pipelines requires re-validation against regulatory air quality standards and retraining with local aerosol climatologies to avoid degradation in new geographic regions.
Limitations and When Not to Use This
The paper assumes that satellite-observable spectral signatures contain sufficient information to infer AOD without explicit meteorological inputs (surface reflectance, water vapor, ozone)—but in reality, clouds, heavy aerosol loading, and surface changes can confound this relationship, potentially causing the model to fail on out-of-distribution cases. ViTCG inherits the foundation model's sensitivity to domain shift: a model trained on data from summer 2023 may perform poorly on winter imagery or new sensor instruments with different spectral calibrations, requiring continuous retraining or adaptation. The paper does not address temporal consistency—AOD should evolve smoothly over days, but a frame-by-frame transformer may produce unrealistic temporal jitter; integrating temporal context (recurrent or video models) remains open. Computational cost at inference time on continental-scale satellite data streams is not quantified; if latency exceeds near-real-time requirements (~hours), adoption in operational forecasting will be blocked. Finally, the method likely requires orders of magnitude more labeled data than physics-based approaches to achieve comparable accuracy, which may be infeasible in data-sparse regions or for historical reanalysis.
Research Context
This work builds on the recent success of Vision Transformers (ViT) and foundation models in computer vision, adapting them to geophysics and Earth observation—a domain where physics-informed and hybrid approaches have historically dominated. Prior data-driven AOD retrieval methods (using CNNs or shallow networks) treated pixels independently or used local spatial context only, losing information about regional-scale aerosol patterns; ViTCG's long-range attention addresses this limitation. The paper contributes to a growing research direction of applying foundation models to scientific imagery (medical imaging, remote sensing, climate modeling), demonstrating that pre-training on large unlabeled satellite data followed by fine-tuning on small labeled AOD benchmarks can be competitive with domain-specific methods. This opens the door to future work on multi-task learning (joint AOD, cloud, surface reflectance prediction), transfer learning across sensors (training on Sentinel-5P, inference on TROPOMI), and physics-informed regularization (encoding radiative transfer constraints as loss penalties to stabilize learning).
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::