Methane Plume-Artifact Classification

• The paper presents a scheme integrating matched filter enhancement, CNN segmentation, and domain adaptation to accurately differentiate methane plumes from sensor artifacts.
• It employs a tile-level classification strategy with physically motivated retrieval products and data normalization to manage noise across platforms.
• Key performance gains include reduced false positives and improved plume localization through feature-guided spectral analysis and post-classification filtering.

Methane plume-artifact classification is the remote-sensing task of deciding whether an apparent methane enhancement corresponds to a true plume or to a non-plume structure produced by noise, clouds, surface reflectance, retrieval residuals, or sensor artifacts. In current literature, the problem appears both as false-positive filtering on top of matched-filter enhancement products and as binary segmentation of plume versus background; in both cases, the central objective is the same: preserve physically plausible methane plumes while rejecting artifact-driven detections across heterogeneous scenes and across sensors whose spatial resolution, retrieval physics, and noise statistics differ substantially (Mancoridis et al., 2 Jun 2025, Tiemann et al., 2024).

1. Problem formulation and artifact taxonomy

In the tile-classification formulation used for AVIRIS-NG and EMIT, the input is a remote-sensing product derived from a hyperspectral imaging spectrometer, and the output is a binary decision at the tile level: “plume” if the tile contains at least one anthropogenic methane plume, or “background / non-plume” otherwise. This is explicitly a false-positive filtering problem on top of a matched filter detector: matched-filter maps highlight enhancements that may be methane plumes or artifacts, and CNNs are trained to distinguish them (Mancoridis et al., 2 Jun 2025).

A true methane plume is described as a localized, coherent spatial structure in the enhancement map, typically aligned with wind or facility geometry, with elevated methane enhancement relative to background and often an elongated morphology. Artifacts or false positives include cloud contamination, thin cirrus, high-albedo linear features such as roads, terrain and surface spectral effects, retrieval noise, and false “linear plumes” that can arise from simplified retrievals. In EMIT, cloud contamination can inflate variance and mimic plume-like enhancements; in AVIRIS-NG, false linear plumes can arise from the retrieval itself (Mancoridis et al., 2 Jun 2025).

The same distinction appears in segmentation papers, even when “artifact classification” is not the nominal task. In EMIT segmentation work, any pixel wrongly labeled as plume is effectively an artifact, and any pixel wrongly labeled as background is a missed plume. In AVIRIS-NG spectral-spatial detection, artifacts are explicitly associated with roofs, concrete, bright soils, water, vegetation, haze, changing solar elevation, band noise, striping, misregistration, and residuals from imperfect background covariance modeling (Quintero et al., 24 Jun 2026, Kumar et al., 2023).

A recurring theme in the broader survey literature is that plume-artifact confusion is sensor-dependent. Coarse products are affected by sub-pixel heterogeneity and cloud contamination, whereas higher-resolution hyperspectral systems additionally expose striping, spectral confusers, and retrieval-induced false enhancements. This suggests that plume-artifact classification is not merely a generic image-recognition problem; it is a sensor- and product-conditioned discrimination problem grounded in radiative transfer, background modeling, and plume morphology (Tiemann et al., 2024).

2. Data representations, preprocessing, and labels

A defining characteristic of recent classifiers is the use of physically motivated retrieval products rather than raw imagery alone. In the AVIRIS-NG to EMIT domain-adaptation study, the only classifier input is a single-channel Column-wise Matched Filter tile of shape $256\times256$ pixels. Images are tiled into non-overlapping $256\times256$ patches; positive tiles are centered on plumes; train, validation, and test tiles are geospatially separated; negative CMF values are removed; and each dataset is clipped at 0 and its 95th percentile, then normalized to $[0,1]$ by an instrument-specific 95th-percentile scale. For EMIT, cloud filtering rejects tiles with cloud fraction $\ge 20\%$, leaving 327 positive tiles and 8,664 negative tiles in the cloudless set. The airborne AVIRIS-NG corpus contains about 2,000 labeled plume tiles and 25,000 background tiles (Mancoridis et al., 2 Jun 2025).

This tile-level regime contrasts with pixel-level segmentation corpora. The MHS dataset introduced for MethaneMapper contains about 1,185 AVIRIS-NG flightlines, about 3,961 plume sites, concentration maps in ppm, segmentation masks, and derived bounding boxes. The MPDataset used for EMIT multimodal segmentation contains 4,172 samples of size $512\times512$, at about 60 m/pixel, with RGB synthesized from EMIT radiance, methane enhancement from EMITL2BCH4ENH, and plume masks from EMITL2BCH4PLM (Kumar et al., 2023, Quintero et al., 24 Jun 2026).

Across these datasets, explicit artifact labels are uncommon. Most corpora collapse all non-plume structure into a background or non-plume class. In MHS, every pixel not covered by a plume mask is implicitly negative, and therefore includes terrain, atmospheric variability, and spectrally confusing surfaces. In MPDataset, all non-plume structures are background, even though the enhancement map may contain strong but spurious enhancement from artifacts. This means that plume-artifact classification is typically learned as plume-versus-background discrimination with artifact examples embedded inside the background class rather than as a three-way taxonomy (Kumar et al., 2023, Quintero et al., 24 Jun 2026).

The same principle appears in multispectral work. In the PRISMA framework based on synthetic Sentinel-2-to-PRISMA transfer, synthetic plumes are inserted into real PRISMA backgrounds that already contain clouds, roads, buildings, terrain, and retrieval artifacts. The negative class is therefore broad and realistic, even without a dedicated artifact label (Groshenry et al., 2022).

3. Algorithmic paradigms for plume-artifact discrimination

A common baseline for tile-level classification is GoogLeNetAA, an anti-aliased Inception architecture with about 13.4 million learnable parameters. It takes normalized single-channel $256\times256$ CMF tiles, is trained with binary classification at learning rate $1\times10^{-4}$, batch size 16, and 100 epochs, and is intended to capture multi-scale plume morphology while reducing sensitivity to small shifts or pixel misalignments. In the AVIRIS-NG/EMIT study, separate GoogLeNetAA models are trained for AVIRIS-NG campaigns and for EMIT cloudless imagery (Mancoridis et al., 2 Jun 2025).

Hyperspectral end-to-end models push the discrimination problem deeper into spectral space. MethaneMapper combines two ResNet-50 backbones, a transformer encoder-decoder, and two methane-specific modules: a Spectral Feature Generator and a Query Refiner. Its Spectral Linear Filter replaces a single global background covariance with class-specific covariance matrices derived from NDVI/NDWI-based land-cover classes: $$\mathrm{SLF}(x_i) = \frac{(x_i-\mu_k)^T \mathrm{Cov}_k^{-1} t}{\sqrt{t^T \mathrm{Cov}_k^{-1} t}}.$$ The resulting methane-candidate map is used to refine transformer queries toward methane-like regions and away from confusers. Ablations show that DETR without SFG/QR reaches about $mAP\sim0.44$ and $mIoU\sim0.59$ on MHS, whereas adding SFG with a strong feature extractor and QR reaches about $256\times256$0 and $256\times256$1; the paper attributes much of that gain to fewer false positives and better localization (Kumar et al., 2023).

Multimodal segmentation on EMIT introduces a related but distinct strategy: feature-guided methane enhancement. In this model, RGB semantics from a DINOv3 ViT-S/16 branch are modulated by methane-enhancement features from a ResNet-18 branch using

$256\times256$2

This gate amplifies RGB features where methane cues are strong and attenuates them where they are weak or ambiguous. With Dice plus focal loss, the method reaches $256\times256$3 MIoU, $256\times256$4 mean precision, and $256\times256$5 recall on MPDataset, improving over MPSUNet by $256\times256$6 MIoU, $256\times256$7 mean precision, and $256\times256$8 recall. Because higher precision directly implies fewer false positives, the paper interprets these gains as improved artifact rejection (Quintero et al., 24 Jun 2026).

Other systems use explicit instance-level post-classification. In MethaneSAT, Mask R-CNN with a ResNet-50 backbone is followed by a morphology-based filter and a distribution-based classifier using methane and albedo statistics, including Quantile Normality Deviation descriptors. This produces two operating modes: a high-sensitivity mode and a high-precision mode, allowing explicit control over the trade-off between missed plumes and artifact suppression (Pérez-Carrasco et al., 22 May 2026).

4. Cross-platform alignment and domain adaptation

Cross-platform plume-artifact classification is difficult because matched-filter distributions shift strongly across instruments. Between AVIRIS-NG and EMIT, the shift arises from sensor altitude and resolution, sub-pixel mixing, retrieval algorithms, and different noise and artifact characteristics; EMIT, in particular, exhibits strong false enhancements from clouds and operates at about 60 m GSD versus about 3–7 m for AVIRIS-NG (Mancoridis et al., 2 Jun 2025).

One response is transfer learning. Starting from the best airborne AVIRIS-NG GoogLeNetAA model, the EMIT study fine-tunes on EMIT cloudless data with different layer-freezing strategies. The best result is obtained by unfreezing 9 deeper layers and leaving the top 3 layers frozen, training for 20 epochs. On EMIT, this yields precision $256\times256$9, recall $[0,1]$0, and $[0,1]$1, compared with $[0,1]$2 for an EMIT-only model and $[0,1]$3 for zero-shot application of the airborne model. Even with 25% of the EMIT training data, most of the gain over zero-shot transfer is recovered (Mancoridis et al., 2 Jun 2025).

A more aggressive response is unsupervised domain translation. The same study trains a balanced CycleGAN with generators $[0,1]$4 and $[0,1]$5, together with cycle consistency: $[0,1]$6 The best model uses the vanilla GAN objective, $[0,1]$7 cycle consistency, and learning rate $[0,1]$8. Operationally, EMIT CMF tiles are translated to an AVIRIS-like domain and then classified by the airborne classifier. This yields the best reported cross-platform result: for translated EMIT test tiles, the positive class reaches $[0,1]$9, precision $\ge 20\%$0, and recall $\ge 20\%$1, while the negative class reaches $\ge 20\%$2, precision $\ge 20\%$3, and recall $\ge 20\%$4 (Mancoridis et al., 2 Jun 2025).

An important limitation is explicit in the same paper: CycleGAN does not learn to suppress false enhancements per se; it reproduces domain statistics. In some cases, such as a high-albedo road causing a linear false enhancement in EMIT, the translated AVIRIS-like output enhances the road rather than suppressing it. The downstream CNN must still classify the translated feature as non-plume from its global spatial context. This directly counters the common misconception that domain translation alone eliminates artifacts (Mancoridis et al., 2 Jun 2025).

Sensor harmonization can also be embedded in the retrieval stage rather than in the classifier. HyGAS and its Tanager-1 extension use column-wise CMF or CWCMF, in which background statistics are estimated per detector column to reduce structured false positives caused by pushbroom non-uniformities and striping. This is especially important for Tanager-1, which combines high SNR in the methane window with the most pronounced column-dependent variability among Tanager-1, EnMAP, and PRISMA (Ferrari et al., 9 May 2026, Ferrari et al., 9 May 2026).

5. Performance patterns, operating regimes, and error modes

Quantitatively, domain shift causes large degradations if left untreated. In the AVIRIS-NG/EMIT study, the standalone three-campaign AVIRIS-NG classifier reaches $\ge 20\%$5 on AVIRIS-NG but only $\ge 20\%$6 on EMIT, while the standalone EMIT model reaches $\ge 20\%$7 on EMIT and $\ge 20\%$8 on AVIRIS-NG. The CycleGAN-plus-airborne-classifier pipeline therefore represents a large shift in the precision-recall regime: when it predicts “plume” on translated EMIT data, precision is $\ge 20\%$9, indicating very strong false-positive filtering, even though plume recall remains lower at $512\times512$0 (Mancoridis et al., 2 Jun 2025).

The same precision-oriented pattern appears in MethaneSAT. Fine-tuning a Mask R-CNN from MethaneAIR yields scene-level precision $512\times512$1 and recall $512\times512$2 at the baseline operating point. After physics-informed post-processing, the high-sensitivity mode reaches precision $512\times512$3 and recall $512\times512$4, while the high-precision mode, which adds the QND classifier, reaches precision $512\times512$5 and recall $512\times512$6. The paper explicitly interprets the latter as an operating mode for confident source attribution, where artifact suppression is prioritized over exhaustive recall (Pérez-Carrasco et al., 22 May 2026).

Multispectral Sentinel-2 systems display a similar trade-off. On the 2024 held-out test set for MARS-S2L, operating at threshold 0.5 gives recall about $512\times512$7, precision about $512\times512$8, and FPR about $512\times512$9; at the same threshold, a simple MBMP baseline yields recall $256\times256$0 but FPR $256\times256$1. This means the learned model preserves similar recall while reducing false positives by roughly an order of magnitude relative to MBMP thresholding, which the paper treats as artifact-rich (Allen et al., 26 Nov 2025).

In fully automatic EMIT processing, the addition of spectral fit scoring to ML morphology sharply alters review efficiency. When candidates are prioritized using the spectral fit score rather than ML confidence alone, the system yields 63% fewer false detections after 100 inspected predictions, and at 1,000 inspected predictions it yields 71% more detected events and 46% more total plume mass. In daily-digest mode, using stringent spectral thresholds, the deployed system reports negligible false positives (Růžička et al., 5 May 2026).

Across these studies, the hardest residual errors are consistent. Positive plumes remain harder than negatives because plume appearance varies strongly with emission rate, wind, and background. Small or weak plumes, diffuse tails, cloud-contaminated scenes, and artifact patterns that borrow plume-like geometry—especially roads, striping, and cloud-edge structures—remain the principal failure modes. This suggests that plume-artifact classification is fundamentally constrained by the overlap between physical plume morphology and artifact morphology, not only by classifier capacity (Mancoridis et al., 2 Jun 2025, Pérez-Carrasco et al., 22 May 2026).

6. Operational systems, uncertainty, and unresolved issues

Operational methane systems increasingly treat artifact rejection as a joint problem of retrieval design, segmentation, and uncertainty propagation. HyGAS standardizes methane processing across PRISMA, EnMAP, Tanager-1, EMIT, and GHGSat, with matched-filter variants, spectrally matched background selection, scale-aware segmentation, and explicit decomposition of uncertainty from instrument noise and scene-driven clutter through enhancement maps, IME, and emission-rate inversion. The framework’s central practical point is that structured false positives, especially pushbroom artifacts and clutter, must be modeled before flux quantification; otherwise artifact-driven masks bias IME and flux (Ferrari et al., 9 May 2026).

The Tanager-1 comparison sharpens that point. Tanager-1 shows the best radiance-normalized SNR in the methane window after normalization, but also the strongest striping amplitude. EnMAP exhibits the lowest residual striping and smoother enhancement backgrounds. The practical implication is that high SNR does not guarantee low artifact rates; column-dependent radiometric variability can dominate the apparent plume field unless CWCMF or an equivalent column-aware background model is used (Ferrari et al., 9 May 2026).

End-to-end radiance models reinforce the same conclusion from a different direction. MAPL-EMIT uses full EMIT radiances, geometry metadata, and a Swin-v2-S encoder with multi-task outputs for enhancement, plume masks, and source locations. The system supplements ML predictions with model-generated spectral fit scores and estimated noise levels to further limit false-positive rates. On real-world benchmarks, it captures 79% of known hand-annotated NASA L2B plume complexes across 1,084 EMIT granules, captures twice as many plausible plumes as identified by human analysts, and after land/water filtering and spectral vetting yields an upper bound of about 0.16 false plumes per $256\times256$2 km$256\times256$3 granule (Batchu et al., 11 Apr 2026).

Another unresolved issue is representation bias induced by missing pixels. In TROPOMI methane plume detection, low-coverage images can cause a spurious association between coverage and label, leading models to under-detect plumes when cloud-driven missingness is high. The paper on missing-pixel bias reports that both imputation and a weighted resampling scheme that balances classes within coverage bins can significantly reduce this representation bias without hurting balanced accuracy, precision, or recall, and that the debiased models have a higher chance of detecting plumes in low-coverage images (Wąsala et al., 22 Oct 2025).

A final misconception addressed across the literature is that artifact rejection can be reduced to a single threshold on an enhancement map. Recent work instead points toward a layered interpretation: physically motivated matched filters to generate candidate enhancement, morphology-aware neural segmentation to detect plume-shaped structure, spectral or uncertainty-based vetting to confirm gas identity, and domain or sensor adaptation to absorb platform-specific artifacts. A plausible implication is that methane plume-artifact classification has become a unifying problem across airborne, spaceborne, multispectral, and hyperspectral methane monitoring rather than a narrow post-processing step on any one instrument (Mancoridis et al., 2 Jun 2025, Tiemann et al., 2024).