MetaVIn Project - Atmospheric Turbulence Estimation

Accurate Turbulence Strength Without Expensive Equipment

A new approach to long-range image understanding under atmospheric turbulence.
WACV 2025 Paper ID #724
Authors: Ripon Kumar Saha (ASU), Scott McCloskey (Kitware), Suren Jayasuriya (ASU)

WACV 2025 Paper ArXiv Preprint Code

Project Overview

Atmospheric turbulence introduces blur and geometric distortion in long-range images by bending light rays through varying refractive indices. Its severity is expressed by the refractive index structure parameter C_n². Traditional ways to measure C_n² need complex and expensive optical gear, limiting broad adoption.

Our solution: MetaVIn is a Meteorological and Visual Integration system that combines image sharpness metrics with weather station data in a Kolmogorov Arnold Network (KAN). This enables accessible and accurate turbulence strength estimation, vital for improving remote sensing, UAV surveillance, astrophotography, and more.

Crops illustrating turbulence levels — Single frames with varying turbulence intensities. MetaVIn accurately infers `C_n²` by fusing meteorological and visual features.

The Challenge

For computer vision tasks analyzing distant objects, atmospheric turbulence often degrades images with random blur, distortions, and jitters. These effects worsen with:

Large propagation distances
Extremes in temperature, humidity, or wind
Vibrations from camera motion
Changing sunlight or solar loading

Estimating C_n² helps quantify this degradation. But existing methods often use:

Scintillometers or wavefront sensors (too expensive & bulky)
Video-based approaches that fail in real-world conditions with motion
Meteorological data alone that can’t capture path-averaged turbulence
Generic blind IQA metrics poorly correlated to actual turbulence

We offer a simpler method using simple sharpness measures from a single frame, combined with affordable weather station data.

Our Approach

Visual & Meteorological Fusion

MetaVIn extracts sharpness metrics (sum of Laplacian, Tenengrad, variance of gradients) from a single video frame to gauge blur. Simultaneously, it gathers meteorological data (temperature, wind, pressure, humidity, solar loading) plus a single distance measurement from a laser rangefinder to the scene.

Kolmogorov Arnold Network (KAN)

KAN Architecture — KAN architecture fusing meteorological signals, image sharpness, and distance to predict C_n².

We feed these inputs (3 image features, 9 weather metrics, plus distance) into a Kolmogorov Arnold Network. This specialized architecture uses learnable univariate activation functions to capture non-linear relationships between environment, image quality, and turbulence strength.

With just a few hidden units and a high learning rate, the KAN efficiently predicts the log-scaled C_n². This avoids the need for huge training sets or complex computing hardware.

Dataset & Setup

We test on a large dataset of 35,364 samples across multiple geographic sites. Each sample has ground-truth C_n² from a large-aperture scintillometer, plus weather station readings (temperature, wind speed, humidity, solar loading, etc.) and single-frame images from multiple videos. A cost-effective laser rangefinder provides distance to the scene.

If you would like to use the dataset titled “Expanding accurate person recognition to new altitudes and ranges: The briar dataset” by David Cornett, please contact the first author.

Dataset	Location	Date	Samples (BRS/BTS)
BRS1.1	ORNL, TN	Nov 2021	1663 / —
BRS2	Perry, GA	Mar–Apr 2022	9083 / —
BRS3 / BTS3	ORNL, TN	Aug–Sep 2022	8498 / 4305
BRS4 / BTS4	Glen Ellyn, IL	Jan 2023	7751 / 4064

We train on some splits (BRS datasets) and test on others (BTS datasets), ensuring robust evaluation across diverse weather conditions (temperatures from -5.3°C to 32.8°C, wind up to 19 m/s, solar loading up to 1223 W/m²). We also perform data imputation to fill occasional gaps in sensor readings.

Key Results

MetaVIn significantly outperforms standard blind IQA metrics and purely image-based deep learning approaches. Our method achieves:

Spearman correlation: 0.943 (best overall)
Low mean absolute error (MAE) on log-scale C_n²
Robustness to vibration, moving objects, and real-world weather

By leveraging synergy between atmospheric conditions and image clarity, we see better predictions than relying on any single modality. Below is a quick comparison table summarizing performance:

Method	Spearman ↑	MAE ↓	Rel. Error ↓
Classical IQA (BRISQUE, NIQE, etc.)	≤ 0.14	≥ 0.78	≥ 0.057
Gradient-based Passive	0.079	0.631	0.079
Deep CNN (EfficientNetV2)	0.762	0.354	0.025
MetaVIn (This Work)	0.943	0.177	0.006

Scatter of predicted vs. ground truth — Comparisons of predicted vs. ground-truth C_n² across setups. MetaVIn’s results align closely with true values across broad conditions.

SHAP analysis — Feature importance analysis underscores synergy between weather metrics + image sharpness + distance for robust turbulence estimation.

Project Demo Video

Watch a short walkthrough of our approach, including how MetaVIn fuses meteorological data and image features:

Contact & Future Plans

For details on using or extending MetaVIn, or for collaborations, please get in touch with us.

Limitations: MetaVIn currently uses only single-frame metrics (no full spatiotemporal analysis). We also rely on co-located weather stations and a single laser rangefinder measurement.

Future Work: We plan to integrate spatiotemporal features for continuous C_n² tracking, optimize real-time GPU pipelines, and adapt to diverse climates. Our goal is pushing the frontier of meteorological + imaging synergy for robust long-range vision.

Reach us via the WACV 2025 paper authors or through GitHub (#).