From Snapshots to Fluxes: Independent Wind Retrieval Algorithms for Next-Generation Multi-image Greenhouse Gas Satellites via Deep Learning

As the international community moves towards the second Global Stocktake under the Paris Agreement, the demand for independent, transparent, and verifiable greenhouse gas (GHG) emission estimates has never been more critical. While satellite-based monitoring offers a powerful verification tool, the uncertainty of top-down flux estimates is currently dominated by substantial uncertainties in local wind speed input, which typically relies on coarse meteorological reanalysis models. This dependency introduces potential biases and correlated errors that undermine the scientific integrity required for high-stakes climate policy. Addressing this bottleneck, we present a comprehensive science study dedicated to developing an independent, data-driven in-plume wind retrieval framework designed specifically for potential future satellite missions equipped with multi-angle or multi-platform observations. By simulating the data products of such missions using high-resolution Large-Eddy Simulations (LES), we generated a robust dataset of realistic plume dynamics to develop and validate our algorithms. Exploring the temporal information embedded across consecutive plume images, we propose and evaluate two distinct, complementary methodologies for deriving wind velocity fields directly from plume imagery. First, we apply an Multi-Image Correlation Image Velocimetry (CIV) algorithm, optimized to dynamically correct temporal centering errors by averaging correlation surfaces across the observations sequence. Second, we introduce CVision-CIV, a novel deep learning approach based on the UnLiteFlowNet-PIV architecture, which utilizes Convolutional Neural Networks (CNNs) to extract morphological flow features directly from noisy imagery sets. Through an application on simulated CO2 emission plumes, we  demonstrate that while physical CIV methods provide robust baselines, the CVision-CIV model exhibits superior stability in low signal-to-noise regimes, effectively suppressing sensor noise where traditional correlation breaks down. By validating these parallel pathways on LES-generated observations, this work establishes a comprehensive algorithmic foundation for defining observational requirements for future missions aiming to replace reanalysis proxies with precise, observation-based wind products for improved GHG monitoring. We will discuss the methods’ sensitivity to observational noise, number of images, time-difference between images and resolution.