Abstract
Atmospheric turbulence significantly degrades long-range imaging, complicating tasks in surveillance and remote sensing. Traditional measurement methods like scintillometers are prohibitively expensive, while passive vision-only methods often lack robustness in uncontrolled environments. This paper introduces MetaVIn, a multimodal framework that estimates the refractive index structure parameter ($C_n^2$) by fusing single-frame visual sharpness metrics with local meteorological data and path distance. Utilizing a Kolmogorov-Arnold Network (KAN), MetaVIn achieves a Spearman correlation of 0.943 with ground truth data, significantly outperforming existing deep learning and vision-based baselines.
Introduction
Long-range imaging is inherently limited by atmospheric turbulence, which causes geometric distortion and blurring due to refractive index fluctuations. Quantifying this turbulence via the $C_n^2$ parameter is critical for image restoration. However, high-fidelity sensors like scintillometers are costly and difficult to deploy. Conversely, passive video-based estimation often struggles with non-turbulent motion (e.g., camera shake or moving objects) and fails to account for the physical drivers of turbulence, such as solar loading and temperature gradients.
Methodology
MetaVIn addresses these limitations by integrating visual, meteorological, and geometric data. The framework processes three distinct input streams:
- Visual Features: Laplacian-based sharpness, Tenenbaum Gradient, and Variance of Gradient are extracted from a single frame to capture high-frequency degradation.
- Meteorological Data: Nine parameters (e.g., temperature, solar loading, humidity) provide the physical context for turbulence formation.
- Distance: A laser rangefinder measurement provides the path length, essential for calculating path-averaged $C_n^2$.
The core architecture employs a Kolmogorov-Arnold Network (KAN). Unlike standard Multi-Layer Perceptrons (MLPs), KANs utilize learnable activation functions on edges rather than fixed ones at nodes. This allows the model to more effectively approximate the complex, non-linear physical relationships inherent in atmospheric optics.
Experimental Setup
The model was trained and validated on the BRIAR dataset, comprising 35,364 samples collected across diverse geographic locations (Tennessee, Georgia, Illinois) and weather conditions. This represents the most extensive study of its kind, ensuring high generalizability across "wild" environments.
Results and Discussion
MetaVIn demonstrates superior performance compared to both traditional and deep learning baselines. Ablation studies highlight the necessity of multimodal integration: vision-only models struggle with path-dependency, while weather-only models fail to capture local variations along the optical path.
Key findings include:
- High Correlation: MetaVIn achieved a Spearman correlation of 0.943, far exceeding the 0.762 achieved by EfficientNetV2 and the 0.079 of traditional gradient-based methods.
- Model Efficiency: The KAN architecture outperformed standard Fully Connected Neural Networks (FCNNs) in both accuracy and parameter efficiency for this regression task.
- Robustness: The system remains accurate in dynamic scenes with moving subjects, a common failure point for temporal-based turbulence estimation methods.
- Failure of Standard IQA: The study confirmed that standard Blind Image Quality Assessment (BIQA) tools like BRISQUE correlate poorly with turbulence, as they are not designed for the specific physics of atmospheric blur.
Conclusion
MetaVIn provides a cost-effective and highly accurate alternative to specialized hardware for atmospheric turbulence estimation. By fusing visual cues with physical meteorological data through a KAN architecture, the framework enables robust $C_n^2$ prediction in real-world environments. Future work will focus on incorporating temporal video features and optimizing the pipeline for real-time restoration applications.
Summary updated