Di, Xu, An NDMI-Guided Bi-Temporal Network for Methane Plume Detection and Segmentation from Sentinel-2 Observations
Abstract: Methane (CH₄) is a potent greenhouse gas requiring accurate monitoring to support mitigation efforts. While satellite-based observation offers a unique opportunity to monitor methane globally and identify large emission sources, current systems face trade-offs in spatial, spectral, and temporal resolution. Sentinel-2, with its fine spatial resolution and open accessibility, provides valuable potential for large-scale methane detection, but distinguishing true plumes from background signals remains challenging. In this talk, I will introduce a deep learning framework designed to detect and segment methane plumes from Sentinel-2 imagery. The approach integrates temporal and spectral information to improve plume identification across diverse land surfaces. I will also share key results from large-scale evaluations and case studies, highlighting how this method advances automated, scalable methane monitoring from space.
Bio: Di Xu is a PhD student in the Department of Earth Science and Engineering at Imperial College London. Her recent work focuses on large-scale and long-term monitoring of methane point-source emissions using satellite observations. Her research interests include remote sensing, environmental monitoring, and the development of data-driven methods for greenhouse gas detection and quantification.
Judy, Zhou, Score-Based Generative Modeling with Variational Data Assimilation–Inspired Conditioning for Real-time Satellite CO Reconstruction 

Abstract: The TROPOspheric Monitoring Instrument (TROPOMI), onboard Sentinel-5 Precursor, provides daily global measurements of total column CO (TCCO) at unprecedented spatial resolution. However, the real-time TROPOMI product often contains large data gaps due to cloud cover, high solar zenith angles, and unfavorable surface conditions, limiting its use in downstream applications. This study proposes a Score-Based Generative Modeling with Variational Data Assimilation–Inspired Conditioning (Var-SGM) for reconstructing seamless, high-resolution TCCO fields directly from incomplete Level-2 retrievals. The model is built on a score-based generative framework that learns the distribution of TCCO as a prior physical knowledge of atmospheric states. During generation, a conditioning strategy inspired by variational data assimilation is applied to ensure observational consistency while accounting for measurement uncertainty, correcting biases, and preserving spatial coherence. Unlike many machine learning approaches, Var-SGM requires no auxiliary predictors or reanalysis inputs, enabling real-time operation. Evaluation against independent in-situ observations, a reanalysis dataset, a state-of-the-art deep learning–based multi-data fusion product, and retrievals from the Infrared Atmospheric Sounding Interferometer (IASI) demonstrates that the Var-SGM outperforms geostatistical interpolation baselines, achieves accuracy comparable to advanced multi-source fusion methods, and recovers fine-scale features beyond those resolved in reanalysis, and produces spatial patterns consistent with IASI observations. The framework reconstructs daily continental-scale fields within approximately two seconds on a single GPU, highlighting the potential of Var-SGM as an efficient, self-contained solution for atmospheric remote sensing gap-filling in downstream real-time applications. 

Bio: Judongyang (Judy) Zhou is a third-year PhD student in the Department of Earth Science and Engineering at Imperial College London. Her research focuses on integrating deep learning with data assimilation to enhance both satellite observation reconstruction and atmospheric composition forecasting. Her work aims to advance the use of satellite and modelling systems for better understanding, monitoring, and predicting air quality and climate. 

Add event to calendar
See all events