Google 与 NASA JPL 发布 MAPL-EMIT 深度学习框架，用卫星高光谱数据绘制全球甲烷排放

谷歌研究院研究工程师Vishal Batchu和研究科学家Michelangelo Conserva。

甲烷分析与羽流定位与EMIT模型是一种深度学习框架，能够自动化全球范围内甲烷羽流的检测、增强量化和源估计，将原始卫星数据转化为可扩展的气候行动。

甲烷：https：//en.wikipedia.org/wiki/甲烷是一种强效温室气体;在100年的时间范围内，其变暖潜力是二氧化碳的30倍：https：//www.ipcc.ch/assessment-report/ar6/。事实上，自工业时代开始以来，甲烷推动了约25%的人为变暖：https：//www.ipcc.ch/assessment-report/ar6/。由于甲烷的大气寿命相对较短，及时减少这些排放为缓解全球气温上升提供了关键的“快速行动”途径。

这一紧迫性体现在全球甲烷承诺：https：//www.globalmethanepledge.org/ 中，已有125多个国家承诺到2030年减少30%的排放。为实现这些目标，我们必须赋能利益相关者追踪废弃物、农业和能源领域的局部点源（仅数十米空间足迹产生的排放）。最具成本效益的策略是减少石油和天然气基础设施、农业设施和垃圾填埋场的排放。

为了在全球范围内追踪这些排放，科学家们越来越依赖基于空间的成像技术。一个典型例子是美国宇航局（NASA）国际空间站上的地表矿物尘埃源调查仪：https：//earth.jpl.nasa.gov/emit/（EMIT）。虽然最初设计用于绘制干旱地区的矿物成分，但NASA喷气推进实验室（JPL）的科学家们利用了EMIT先进的高光谱技术（https：//en.wikipedia.org/wiki/Hyperspectral_imaging 能力来探测甲烷排放。通过记录每个像素数百条不同的光带，研究人员能够“看到”这些原本不可见气体独特的化学指纹。

从太空观察甲烷检测的全球视图。当我们逐渐放大时，MAPL-EMIT 模型会突出显示特定的甲烷羽流，揭示关键细节，例如来源位置。

从太空测量甲烷需要在三个关键因素之间取得平衡：（1）视野（空间覆盖/重访），（2）空间分辨率，以及（3）光谱分辨率。

全球测绘器如 TROPOMI：https://www.tropomi.eu/ 被设计用于通过集成高覆盖范围（约 2,600 公里扫描宽度）、粗空间分辨率（约 5.5 公里 x 3.5 公里）和精细光谱采样（0.1 纳米）来检测背景甲烷浓度的微小变化。

相比之下，点源测绘器如 EMIT 擅长在设施尺度上测量甲烷排放。它们通过将中等覆盖范围（80 公里宽视野）与极高空间分辨率（60 米）和中等光谱分辨率（7.4 纳米光谱采样）结合，实现高信噪比捕捉甲烷的化学特征。

然而，要在全球范围内充分释放这些丰富数据的潜力，还面临额外挑战。地球多样的景观提供了复杂的背景，而一些地表材料可能伪装成甲烷，使得识别较小或较分散的源尤其具有挑战性。为了在 EMIT 团队的基础工作上进一步发展并实现高通量全球测绘，我们与他们合作，应用能够理解场景更广泛视觉上下文的深度学习模型。

这项工作与谷歌在 Google Earth AI：https://ai.google/earth-ai/ 背后的更广泛努力一致，这是我们收集的地理空间模型和数据集，旨在将行星数据转化为可操作的情报。通过大规模将深度学习应用于卫星影像，我们希望用专门工具补充更广泛的行星 AI 项目，用于针对性环境监测。

发现隐形之物。该比较展示了甲烷检测的挑战：标准可见光影像未显示气体，NASA L2B 增强：https://www.earthdata.nasa.gov/data/catalog/lpcloud-emitl2bch4enh-002（基于匹配滤波器）显示噪声信号，但 MAPL-EMIT 模型清晰地将甲烷羽流从背景景观中分离出来。

我们使用端到端视觉变换器架构（Swin-S transformer：https://arxiv.org/abs/2103.14030）构建了 MAPL-EMIT。虽然许多方法是逐像素分析高光谱数据，但 MAPL-EMIT 利用现代计算机视觉技术处理完整的光谱信息及其周围的空间上下文。通过分析气体在景观中的扩散方式，该模型能够更好地区分真正的风吹甲烷羽流（来自特定源的甲烷气体扩散轨迹）和仅仅具有相似光谱特征的地面区域，这在过去常导致甲烷误检。

关键是，这种空间意识使模型能够解开高度复杂的场景。在密集的工业区域，多个相邻设施的排放物经常混合成一个单一的云层。为了理解这些情景，MAPL-EMIT 同时解决三个不同的任务：

解开复杂排放。通过分析空间上下文，MAPL-EMIT 能够同时描绘多重重叠的甲烷羽流的精确形状，并定位其各自的来源（用 X 标记），即使在密集的工业区域也能做到。

基于变换器的模型需要海量数据来学习，但全球范围内包含数百万真实甲烷排放的标注数据集根本不存在。为克服这一问题，我们开发了基于物理的模拟框架。我们创建了 360 万个合成甲烷羽流，并将其直接注入到真实的 EMIT 场景中。通过使用拉格朗日烟羽模型：https://www.nature.com/nature-index/topics/l4/lagrangian-particle-dispersion-modeling-in-atmospheric-studies，该模型模拟了颗粒在空气中的运动和扩散，我们能够重现实际气体排放的混乱和湍流现实。基于这些高度逼真的模拟进行训练，使 MAPL-EMIT 学会在各种大气和地理条件下检测甲烷。这种合成训练方法提供了几个关键优势：

使用合成现实进行训练。由于不存在数百万个带标签的真实甲烷羽流，我们通过将基于物理模拟的羽流（代表不同的排放速率和湍流大气条件）直接注入真实的 EMIT 高光谱场景来训练模型。

在实际卫星数据上的应用表明，MAPL-EMIT 在可扩展排放映射方面显示出强大的潜力。在与 NASA 的黄金标准 L2B 甲烷羽流数据集（https://www.earthdata.nasa.gov/data/catalog/lpcloud-emitl2bch4plm-002）进行基准测试时，该模型捕捉到了 84% 的专家标注羽流，并在约 1100 个 EMIT 数据块中识别出约多 50% 的合理羽流，展示了其从背景噪声中分离微弱信号的能力。更多细节请参见论文：https://www.pnas.org/doi/10.1073/pnas.2612145123。

这种增强的灵敏度还使 MAPL-EMIT 能够可靠地捕捉较弱排放，改进了现有检测极限。该模型在复杂环境中也表现出了稳健性，成功绘制了全球 25 个排放量最高的垃圾填埋场中的 24 个的羽流。

与许多高灵敏度模型一样，假阳性仍然是一个持续的挑战，尤其是在复杂地形中。为帮助缓解这一问题，输出结果与基于物理的羽流置信度（光谱拟合）评分（https://www.sciencedirect.com/science/article/pii/S0034425725002640）配对，根据跨步推理的检测次数进行评估，并使用多种其他属性进行评估，使用户能够根据需要筛选并在捕捉真实羽流能力与假阳性风险之间进行权衡。然而，这并非总是直观的，因此我们还根据这些属性为每个羽流标记“低”或“高”置信度，使用户能够直接使用数据。

追踪持续排放随时间的变化。MAPL-EMIT 甲烷检测的时间序列成功捕捉了来自约旦安曼一处主要垃圾填埋场的羽流，展示了该模型在复杂环境中的高灵敏度和稳健性。

MAPL-EMIT 展示了强大的协作，将谷歌的机器学习专长与 NASA JPL 合作伙伴的领域知识相结合。我们共同推进了设施级空间甲烷观测的全部潜力，为全球社区提供必要工具，从而推动在减少温室气体排放方面采取有意义的行动。

我们要感谢在谷歌和NASA喷气推进实验室（JPL）工作的个人，他们开展了这项工作并使发射成为可能，包括（按字母顺序）：Alex Wilson, Anna M. Michalak, Varun Gulshan, Philip G. Brodrick, Andrew K. Thorpe, Christopher V. Arsdale, Burak Ekim, Carl Elkin, Tal Geller, Omry Gillon, Nita Goyal, Mansi Kansal, Roy Nadler, John Platt, Sergei Shames, Bijoy Shetty, Aaron Sonabend, Deepika Sukhija, Shahar Timnat, Maxim Neumann, Anton Raichuk, Frances Reuland, Adam R. Brandt, David R. Thompson, Robert O. Green, Jay Radzinski, Vishal V. Batchu, 以及 Michelangelo Conserva。

Vishal Batchu, Research Engineer, and Michelangelo Conserva, Research Scientist, Google Research

The Methane Analysis and Plume Localization with EMIT model is a deep-learning framework that automates the detection, enhancement quantification, and source estimation of methane plumes globally, turning raw satellite data into scalable climate action.

Methane：https://en.wikipedia.org/wiki/Methane is a potent greenhouse gas; over a 100-year timeframe, its warming potential is 30 times：https://www.ipcc.ch/assessment-report/ar6/ greater than that of carbon dioxide. In fact, it has driven approximately 25% of human-induced warming：https://www.ipcc.ch/assessment-report/ar6/ since the start of the industrial era. Because methane has a relatively short atmospheric lifespan, promptly reducing these emissions offers a critical "fast-action" pathway to mitigating global temperature rise.

This urgency is reflected in the Global Methane Pledge：https://www.globalmethanepledge.org/, where over 125 countries have committed to a 30% emissions reduction by 2030. To hit these targets, we must empower stakeholders to track localized point sources (emissions occurring from a small spatial footprint on the order of a few tens of meters) across the waste, agriculture, and energy sectors. The most cost-effective strategies are to mitigate emissions from oil and gas infrastructure, agricultural facilities, and landfills.

To track these emissions on a global scale, scientists increasingly rely on space-based imaging. A prime example is NASA’s Earth Surface Mineral Dust Source Investigation：https://earth.jpl.nasa.gov/emit/ (EMIT) instrument on the International Space Station. While originally designed to map mineral composition in arid regions, scientists at NASA’s Jet Propulsion Laboratory：https://www.jpl.nasa.gov/ (JPL) and the broader scientific community have leveraged EMIT's advanced hyperspectral：https://en.wikipedia.org/wiki/Hyperspectral_imaging capabilities to detect methane emissions. By recording hundreds of distinct bands of light for every pixel, it allows researchers to "see" the unique chemical fingerprints of these otherwise invisible gases.

A global view of methane detections from space. As we zoom in, the MAPL-EMIT model highlights specific methane plumes, revealing critical details such as source location.

Measuring methane from space requires balancing three key factors: (1) field of view (spatial coverage/revisit), (2) spatial resolution, and (3) spectral resolution.

Global mappers like TROPOMI：https://www.tropomi.eu/ were designed to detect small changes in background methane concentrations by integrating high coverage (approximately 2,600 km swath width), coarse spatial resolution (around 5.5 km x 3.5 km), and fine spectral sampling (0.1 nm).

In contrast, point source mappers like EMIT excel at measuring methane emissions at the facility scale. They achieve this by combining moderate coverage (an 80 km wide field of view) with very high spatial resolution (60 meters) and a moderate spectral resolution (7.4 nm spectral sampling), sufficient to capture the chemical signature of methane at a high signal to noise ratio.

However, fully unlocking the potential of this rich data at a global scale presents additional challenges. The Earth's varied landscapes provide a complex backdrop, and some surface materials can masquerade as methane, making the identification of smaller or more diffuse sources particularly challenging. To build on the EMIT team's foundational work and enable high-throughput global mapping, we collaborate with them to apply deep-learning models that can understand the broader visual context of the scene.

This work aligns with Google’s broader effort behind Google Earth AI：https://ai.google/earth-ai/, our collection of geospatial models and datasets to turn planetary data into actionable intelligence. By applying deep learning to satellite imagery at scale, we aim to complement broader planetary AI initiatives with specialized tools for targeted environmental monitoring.

Spotting the invisible. This comparison demonstrates the challenge of methane detection: standard visible imagery shows no gas, NASA L2B enhancements ：https://www.earthdata.nasa.gov/data/catalog/lpcloud-emitl2bch4enh-002 (from matched filter) reveal a noisy signal, but the MAPL-EMIT model clearly isolates the methane plume from the background landscape.

We built MAPL-EMIT using an end-to-end vision transformer architecture (Swin-S transformer：https://arxiv.org/abs/2103.14030). While many approaches analyze hyperspectral data on a pixel-by-pixel basis, MAPL-EMIT leverages modern computer vision techniques to process the complete spectrum of light alongside its surrounding spatial context. By analyzing how gas disperses across the landscape, the model is better equipped to distinguish a true, wind-blown methane plume (a trail of methane gas dispersing from a specific source) from a patch of ground that simply shares a similar spectral signature, which has historically caused false methane detections.

Crucially, this spatial awareness empowers the model to untangle highly complex scenes. In dense industrial regions, emissions from multiple neighboring facilities often merge into a single cloud. To make sense of these scenarios, MAPL-EMIT simultaneously solves three distinct tasks:

Disentangling complex emissions. By analyzing spatial context, MAPL-EMIT simultaneously delineates the exact shape of multiple, overlapping methane plumes and pinpoints their respective source origins (marked with an X), even in dense industrial regions.

Transformer-based models require massive amounts of data to learn, but a global, labeled dataset of millions of real-world methane emissions simply doesn't exist. To overcome this, we developed a physics-based simulation framework. We created 3.6 million synthetic methane plumes and injected them directly into real EMIT scenes. By using Lagrangian puff models：https://www.nature.com/nature-index/topics/l4/lagrangian-particle-dispersion-modeling-in-atmospheric-studies, which simulate how particles move and disperse through the air, we were able to recreate the chaotic, turbulent reality of actual gas emissions. Training on these highly realistic simulations allowed MAPL-EMIT to learn to spot methane under a vast variety of atmospheric and geographic conditions. This synthetic training approach provided several key advantages:

Training with synthetic reality. Because millions of labeled, real-world methane plumes do not exist, we trained the model by injecting physics-based simulated plumes, representing diverse emission rates and turbulent atmospheric conditions, directly into real EMIT hyperspectral scenes.

Deployed on real-world satellite data, MAPL-EMIT demonstrates strong potential for scalable emissions mapping. Upon benchmarking against NASA's gold-standard L2B methane plumes dataset：https://www.earthdata.nasa.gov/data/catalog/lpcloud-emitl2bch4plm-002, the model captures 84% of expert-annotated plumes and identifies around 50% more plausible plumes across ~1100 EMIT granules, showcasing its ability to separate subtle signals from background noise. See the paper：https://www.pnas.org/doi/10.1073/pnas.2612145123 for more details.

This increased sensitivity also allows MAPL-EMIT to reliably capture weaker emissions, improving on current detection limits. The model also proved robust in complex environments, successfully mapping plumes at 24 of the world's 25 top-emitting landfills.

As with many highly sensitive models, false positives remain an ongoing challenge, particularly in complex terrain. To help mitigate this, outputs are paired with physics-based plume confidence (spectral fit) scores：https://www.sciencedirect.com/science/article/pii/S0034425725002640, assessed based on the number of detections over strided inference, and evaluated using multiple other properties, enabling users to filter and trade off between the ability to capture real plumes and the risk of false positives as they see fit. However this isn’t always straightforward, which is why we also tag each plume with a “lower” or “higher” confidence based on these properties, allowing users to directly use the data.

Tracking persistent emissions over time. A time-series of MAPL-EMIT methane detections successfully capturing plumes originating from a major landfill in Amman, Jordan, demonstrating the model's high sensitivity and robustness in complex environments.

MAPL-EMIT showcases a powerful collaboration, bringing together Google's machine learning expertise with the domain knowledge of our collaborators at NASA JPL. Together, we are advancing the full potential of space-based methane observations at the facility scale, providing the global community with the tools necessary to enable meaningful action on reducing greenhouse gas emissions.

We would like to thank individuals across Google and NASA JPL who carried out this work and made the launch possible, including (in alphabetical order): Alex Wilson, Anna M. Michalak , Varun Gulshan, Philip G. Brodrick, Andrew K. Thorpe, Christopher V. Arsdale, Burak Ekim, Carl Elkin, Tal Geller, Omry Gillon, Nita Goyal, Mansi Kansal, Roy Nadler, John Platt, Sergei Shames, Bijoy Shetty, Aaron Sonabend, Deepika Sukhija, Shahar Timnat, Maxim Neumann, Anton Raichuk, Frances Reuland, Adam R. Brandt, David R. Thompson, Robert O. Green, Jay Radzinski, Vishal V. Batchu, and Michelangelo Conserva.