Towards Operational Automated Greenhouse Gas Plume Detection

TL;DR

This paper tackles the challenge of operational automated GHG plume detection from imaging spectroscopy by diagnosing core obstacles—data and label quality, spatiotemporal biases, and aligned modeling objectives—and demonstrating that convolutional neural networks can reach operational performance when these issues are mitigated. It introduces a multitask CNN approach that jointly performs instance detection and pixelwise segmentation, and validates it across multicampaign airborne and spaceborne data, emphasizing rigorous tilewise and scenewise evaluation to gauge generalization. The study provides thorough data handling, quality control, and sampling procedures to avoid spatiotemporal leakage, analyzes plume detectability as a function of plume size and concentration, and compares different model architectures (tilewise, pixelwise, multitask) on AVIRIS-NG, GAO, and EMIT datasets. The results highlight the trade-offs between detection and segmentation accuracy, underscore the impact of retrieval differences and data quality on operational readiness, and offer open data, code, and best-practice guidelines aimed at advancing toward deployment-ready automated plume detectors with reproducible validation standards.

Abstract

Operational deployment of a fully automated greenhouse gas (GHG) plume detection system remains an elusive goal for imaging spectroscopy missions, despite recent advances in deep learning approaches. With the dramatic increase in data availability, however, automation continues to increase in importance for natural and anthropogenic emissions monitoring. This work reviews and addresses several key obstacles in the field: data and label quality control, prevention of spatiotemporal biases, and correctly aligned modeling objectives. We demonstrate through rigorous experiments using multicampaign data from airborne and spaceborne instruments that convolutional neural networks (CNNs) are able to achieve operational detection performance when these obstacles are alleviated. We demonstrate that a multitask model that learns both instance detection and pixelwise segmentation simultaneously can successfully lead towards an operational pathway. We evaluate the model's plume detectability across emission source types and regions, identifying thresholds for operational deployment. Finally, we provide analysis-ready data, models, and source code for reproducibility, and work to define a set of best practices and validation standards to facilitate future contributions to the field.

Figures (14)

• Figure 1: Example $\text{CH}_{4}$ plumes with characteristic features of Oil & Natural Gas ($1^\text{st}$ column) infrastructure, Landfills ($2^\text{nd}$ column). Livestock/Manure Management ($3^\text{rd}$ column), Electricity Generation (${4}^\text{th}$ column) and Wastewater Treatment (last column) facilities. $\text{CH}_{4}$ concentration is overlaid as a yellow-red colormap in the range 0--1500 $\text{ppm-m}$ over a desaturated RGB quicklook.
• Figure 2: Left column: example $\text{CH}_{4}$ plumes observed in CMF retrieval images. Center column: Pixelwise agreement between user-provided plume label masks provided by four independent experts. Right column: CMF-guided plume labels with thresholds $\in$ [250,500,1000]$\text{ppm-m}$. Labels were generated using the Multi-Mission Geographic Information System soliman_2025.
• Figure 3: Example common background ($1^\text{st}$ two columns) and false (last three columns) $\text{CH}_{4}$ enhancements. Common background enhancements lack cohesive spatial structure, making them easy to distinguish from plumes."False" enhancements from confuser materials and retrieval artifacts exhibit concentration levels in the range of real plumes, and can be difficult to distinguish from real plumes.
• Figure 4: Components of an operational CNN-based plume detection system driven by GHG retrieval images. A: We are provided a set of GHG retrieval image scenes with pixelwise labels indicating the boundaries of plumes within each scene. B: We employ a spatially stratified sampling and validation strategy to extract tiles from our labeled GHG scenes and partition the extracted tiles by scene into spatially disjoint training and test sets. C: Provided our training/test tiles, our objective is to construct a plume detector that predicts the probability that each pixel in an input GHG image represents a plume. D: Once trained, an operational model will be deployed to detect plumes in full-scene GHG retrieval image products, and we validate candidate plume detectors with respect to scenewise plume instance detection and segmentation metrics accordingly.
• Figure 5: AVIRIS-NG, GAO and EMIT methane plume data sets considered in this work. Data sets IDs used to train/test the airborne (AVIRIS-NG/GAO) methane plume detector model are indicated by blue text, and purple text indicates the data used to train/test the spaceborne (EMIT) methane plume detector. Airborne campaigns used only as hold-out test data for the airborne plume detector are shown in orange text. Red text indicates the data sets with unique characteristics described in detail in Suppl. Sections \ref{['apx:bias_region']} and \ref{['apx:gao_penn']}. These were also only used as hold-out test data.