MethaneSAT (operational March 2024–June 2025) occupied a unique position in the world of methane satellites by providing a comprehensive global picture of oil and gas methane emissions at high spatial resolution. It characterized methane from both dispersed area sources and discrete point sources in more than 80 % of global oil and gas production. With its high spectral resolution (0.24 nm) and spectral sampling (0.08 nm), large swath width (220–440 km) and along-track length (∼ 220 km), and fine spatial sampling (110 m across-track × 400 m along-track at nadir), MethaneSAT provided clear signals of strong point source emissions along with area emissions . In this study, our input data is the full-scene level-3 (L3) methane concentration map (pixel size: 45 × 45 m²) produced by the CO₂ proxy retrieval method, which typically achieves 25–55 ppb precision for a single pixel . Level-2 pixels (110 m × 400 m) with cloud contamination and low-quality data were masked during upstream processing to create the L3 products.

MethaneAIR is an airborne image spectrometer developed as a demonstrator of the MethaneSAT satellite. It is typically flown at 12 km flying altitude, providing a 4.5 km observing swath with fine spatial resolution (5 m across-track × 25 m along-track). MethaneAIR is also designed to provide emissions information from both dispersed area sources and discrete point sources. Previous studies have validated the MethaneAIR performance on point source emissions using different retrieval methods and plume quantification methods in controlled releases . Similar with what we did with MethaneSAT data, we also took the full-scene L3 methane concentration map (pixel size: 10 × 10 m²) produced by the CO₂ proxy retrieval, after cloud and cloud shadow screening at L2.

In this work, we use an application-oriented definition of a “plume” to meet the goals of source detection, localization, and flux quantification from satellite observations. We define a plume as a connected region of enhanced methane concentration that either contains a clear hotspot that enables emission source identification, or exhibits a shape that is physically consistent with transported emissions and allows for source localization.

Under this definition, not all true methane enhancements are classified as plumes. In particular, we exclude cases where emissions may cause real enhancements, but cannot be reliably linked to a source. These include: (1) dispersed small sources whose individual emission rates are below the detection limit; (2) intermittent emissions where the source stops emitting before overpass, resulting in enhancements observed significantly downwind of the source location; and (3) fragmented plumes where turbulence or eddies have broken an extended plume into multiple detached clumps that cannot be reliably traced back to a single source.

Based on this definition, we next describe the processing steps used to identify and mask plumes from the satellite or aircraft XCH₄ observations. The key step is to generate a denoised version of the column-averaged dry-air mole fraction of methane (XCH₄) image by applying a specialized set of 2D wavelet transforms to filter out noise and enhance plumes. Using the denoised image, we then create binary plume masks, determine source locations, and remove false detections. The input XCH₄ image is the full scene of the L3 data from a MethaneAIR flight or a MethaneSAT collection. The detailed steps are as follows.

We focus on denoising retrieved CH₄ concentration maps rather than the full hyperspectral radiance data, as operating on full spectra would require modifying the retrieval pipeline and is beyond the scope of this study. We adapt an approach used in medical image processing not previously implemented in atmospheric studies . In , targets are more spatially coherent and exhibit higher contrast relative to noise than methane concentration maps. Therefore, our method begins with an additional pre-processing step where all pixels above a defined threshold are assigned a uniform value, converting strong signals to low frequencies to help separate them from high-frequency noise. In this step, all pixels above a defined threshold are assigned a uniform value (missing pixels due to upstream processing were filled by statistical values, i.e., mean plus scaling factor times standard deviation). For MethaneAIR, the threshold is uniform across the scene, determined as the mean value of the whole image scene plus a scaling factor (value: 2) times the standard deviation of the whole scene, corresponding to approximately the 97.725th percentile under a Gaussian assumption; and for MethaneSAT, the threshold is adjusted in varying local background, defined as the mean value for a subregion (4.5 × 4.5 km²) of the image plus a scaling factor (value: 1.75) times the standard deviation of this local area, corresponding approximately to the 95.5th percentile under a Gaussian assumption. Then pixels above this threshold are set to the maximum value of the entire image (Fig. a–b). The sensitivity analysis of the scaling factors and the local background variance are discussed in Sect. S1.2 in the Supplement. These parameters were chosen to maximize plume detections while minimizing FPs, and are readily tuned for different observing platforms. Selections should be made to minimize overfitting to the datasets being considered, as discussed further in Sect. . The resulting image is referred to as the “input image” (Fig. b).

After this pre-processing, a multilevel 2D discrete wavelet transform is applied to the input image to generate approximation and detail coefficients. The wavelet filter is applied in both horizontal and vertical directions on an image of size m × n , producing four coefficient matrices of size m2 × n2 (cA1, cH1, cV1, cD1), where cA1 represents the approximation and (cH1, cV1, cD1) represent the details. This decomposition process is repeated recursively on the approximation coefficients, creating additional levels of details, each level capturing finer aspects of the image. The resulting coefficient list includes the approximation coefficients at the Lth level of decomposition and tuples of detail coefficients in descending order of the decomposition level (Eq. ). We used the Haar wavelet in this step for its simplicity and computational efficiency . 1Coefficients=(cAL,cHL,cVL,cDL,…,cH2,cV2,cD2,cH1,cV1,cD1),L=log⁡2min(m,n)/2 After decomposition, a common denoising approach involves setting the detail coefficients to zero and then performing an inverse 2D discrete wavelet transform . This reconstructs an image of the same size as the input, with the approximation retained and high-frequency details removed. However, this approach causes a loss of detailed plume structure, which is essential for plume visualization. To address this, we followed and set the approximation coefficients (cAL) to zero and performed the inverse wavelet transform, building an image that retains only high-frequency components (Fig. c). We then subtract this high-frequency image from the input image, so that the resulting image greatly minimizes noise and enhances plumes while maintaining a detailed plume structure (Fig. d).

In this process, the decomposition level L is a key parameter. Smaller values of L may retain more high-frequency components, resulting in a large number of false positives. In contrast, larger values tend to over-smooth the data and suppress small methane signals, leading to missed detections (Tables S2 and S12 in the Supplement). To balance these effects, we set L at half of its maximum possible value (Eq. ).

Finally, we apply soft thresholding wavelet denoising to the processed image for further denoising. This process begins with a wavelet decomposition, followed by shrinking the wavelet coefficients based on their magnitude, such that the coefficients lower than a defined threshold (usually zero) are reduced to the threshold value, while larger coefficients are preserved . As a non-linear filter, it helps suppress small noise-associated coefficients that persists from earlier processing stages, while retaining larger coefficients likely to represent meaningful features. However, this approach may introduce some degree of over-smoothing, particularly for very weak plumes. Once completed, the wavelet coefficients are transformed back to the original domain, producing the final denoised image (Fig. e). Note that the pixel values in the denoised image have been subjected to non-linear processing and no longer retain their physical meaning and are therefore unsuitable for flux quantification. Instead, the denoised image is used only for plume masking.

We note that for an ideal linear and orthogonal wavelet transform, the key subtraction step of our denoising method would be equivalent to setting the high-frequency coefficients to zero and reconstructing the image. However, this equivalence does not strictly hold in our implementation due to multiple non-linear steps in this framework, including the pre-processing thresholding, decomposition to a finite level, and the subsequent soft-thresholding of wavelet coefficients. These together result in a structure-preserving denoising rather than a true low-pass filtering.

Based on the denoised image, the plume masks are generated in the following two steps. First, pixels above a threshold value are preserved, where the threshold is defined as the local mean plus a scaling factor (value: 1.5 for MethaneAIR, 1.75 for MethaneSAT) times the local standard deviation (the sensitivity tests of these parameters are discussed in Sect. S1.2). Note that the threshold is determined on the basis of the local background in the denoised image to adapt to spatial heterogeneity caused by surface properties, atmospheric conditions, upwind emitters, and retrieval noise. Second, XCH₄ clumps are identified from these preserved pixels using the 8-connectivity connected component algorithm, which groups adjacent pixels into connected components based on their 8 neighboring pixels (“Queen's” definition). Clumps are then preserved if they exceed a size threshold (100 pixels for MethaneAIR, 500 for MethaneSAT; see their sensitivity analysis in Sect. S1.2). Although binary masks are generated from the denoised image, they are applied to the original concentration map for emissions quantification to conserve mass.

Given a set of derived plume masks, we filter out false masks by applying specific criteria based on concentrations, plume shapes, and wind direction.

The concentration filter aims to remove false masks by finding a high concentration “hotspot” in the original concentration map, a smaller XCH₄ clump with elevated pixel values within the mask. The rationale is that a plume typically features a central hotspot with higher enhancements near the source, diffusing into lower enhancements along the wind direction as the plume spreads out.

The hotspot is defined by a pixel value threshold and a size threshold (10 pixels for MethaneAIR, 20 for MethaneSAT, see their sensitivity analysis in Sect. S1.2). The pixel value threshold is adjustable to the regional background concentration variations resulting from different surface types, weather conditions, etc. (see Sect. S1.1 for details on how these values were adjusted for MethaneAIR and MethaneSAT). We define a hotspot filtering criterion named the hotspot ratio, which is the ratio of the hotspots pixel count to the total pixel count of the mask. Masks with a low hotspot ratio (threshold: 0.002) are discarded, while those with a high hotspot ratio (threshold: 0.03) are accepted without further filtering. Masks with a hotspot ratio in between are further evaluated by the latter filters (this is conceptually similar with Hysteresis Thresholding in ).

Plumes can take various shapes affected by factors such as flux rate, duration, wind speed and direction, and instrument detection limits. A well-defined plume may appear long and narrow under high winds or in an ellipse or circular shape under low winds. Sometimes the plume may be split into multiple clumps by local eddies, alternating between regions of higher and lower enhancements, resembling several detached bubble-shaped plumes (Fig. S1 in the Supplement). Therefore, in the shape filtering process, we do not restrict our criteria to “standard” plume shapes. Instead, we discard masks with unusual shapes that are clearly not plume-like.

One specific shape to eliminate is the “spider” shape, which consists of multiple branches extending and curling from the center of the mask (Fig. S2). To quantify the degree of curl in such shapes, we introduce the concept of “fiber length”, defined as the length of the longest path along the mask’s skeleton (Fig. S2). For a mask in spider shape, the curling branches lead to a significantly longer fiber length compared to its major axis length (the longest straight line that lies between two boundary points). Based on this observation, we defined a shape filtering criterion by calculating the ratio of fiber length to major axis length, discarding masks with excessively high ratios. The specific parameters used are provided in Tables S7 and S18.

Plumes typically extend along the wind direction, which provides a useful criterion to filter out false masks. To achieve this, meteorological wind data needs to be collected, and the plume origin needs to be identified. In this study, we use High-Resolution Rapid Refresh (HRRR) wind data for U.S. scenes and Global Forecast System (GFS) wind data for all other international scenes. Based on the wind data, the plume origin is determined as the farthest upwind end of the mask. The plume origin from our algorithm is often useful for attribution of plume emissions to particular facilities.

For scenes captured over multiple hours, typically for airborne campaigns, we obtain a range of hourly wind angles from meteorological datasets for the plume origin. Since meteorological data are often coarse both spatially and temporally, we add an angular buffer (55°) to the upper and lower bounds of the wind angle range (its sensitivity is discussed in Tables S6 and S17). For scenes captured at a specific moment, as is often the case with satellite snapshots, we collect a single wind angle and similarly apply angular buffers to get a wind angle range. Finally, we calculate the plume-derived wind angle based on the plume origin and its major axis. Only the plume whose derived wind angle falls within the meteorological wind angle range is preserved. Note that plumes that passed hotspot filtering are not subjected to wind direction filtering, so that high-volume plumes are not removed even if the wind direction in the meteorological product is wrong.

We note that the wind-direction filtering relies on external meteorological data, which may be too coarsely resolved for applications at finer spatial scales or in complex terrain with higher local flow variability and turbulence influence. This may introduce uncertainty in plume alignment and potentially lead to incorrect filtering. Our current wind-direction buffer is relatively wide; however, mismatches may still occur in some cases. Future work could improve upon this by incorporating plume morphology to infer or refine wind direction independently of external meteorological data.

The divergence integral (DI) growing-box method is used to quantify the flux rate of each plume, applying Gauss's theorem through a series of closed surfaces enclosing the source to estimate the flux divergence at increasing distances from the origin. The DI growing-box method has been introduced in , and here we briefly outline the key steps of implementation. Given a plume origin and the length of the XCH₄ mask, we use the unfiltered XCH₄ map and draw a series of rectangles (“growing boxes”) surrounding the plume origin, with a distance from the origin to the downwind edge of each rectangle ranging from approximately 100 m to the full plume length (up to 10 km). Each successive rectangle is larger than the previous one and completely contains it. The incremental distance between rectangles is one pixel width in each direction, except for the upwind side, where the increment is a quarter of the pixel width. Then based on the measured XCH₄ along each rectangle, we calculate the surface flux divergence integral (kg h⁻¹). The resulting fluxes are then aggregated into a final flux by taking the average over all rectangles that cross the plume. Wind speed and direction from the meteorological product at 80 m height (HRRR or GFS, selected to represent the surface layer windspeed) was used for this calculation, and for sufficiently elongated plumes (eccentricity > 0.87), the wind direction was rotated to match the angle of the major axis of the observed plume.

Although the proposed wavelet-based plume detection operates without manual input, all plume detections presented in the Results section were also manually validated by independent analysts, including attribution to emitting infrastructure through review of high-resolution satellite imagery and cross-checking against plume detections identified by other methods .

We first evaluated the impact of the denoising step by comparing the wavelet denoising step with two alternatives: Gaussian filtering and median filtering. Gaussian filtering used a 2D isotropic Gaussian filter with standard deviation of 3 pixels, and median filtering used a 3 × 3 pixel moving window. We directly replaced the wavelet denoising step with these methods while keeping all subsequent filtering steps identical. The comparison was performed across three MethaneAIR scenes and three MethaneSAT scenes spanning a range of background and wind conditions, consistent with what we used in the sensitivity analysis.

The results (Tables S8 and S19) show that our approach consistently yields a higher number of true detections compared to both Gaussian and median filtering. In addition, the alternative methods produce either comparable or higher numbers of false detections in most of the scenes. This indicates that while conventional denoising techniques can suppress noise, they are less effective at preserving plume structures across spatial scales, leading to reduced detection sensitivity.

Next, we assessed the sensitivity of the results to the choice of wavelet by testing several commonly used families, including Daubechies (db4), Symlets (sym4), and biorthogonal wavelets (bior4.4). These wavelets represent different mathematical properties, such as compact support (db4), increased symmetry (sym4), and linear phase reconstruction (bior4.4) .

As summarized in Tables S9 and S20, we do not observe a significant difference between the Haar wavelet and the alternative wavelets in terms of the number of true and false detections. In most cases, the other wavelets yield equal or fewer true detections and equal or more false detections. Furthermore, the plume masks generated using the Haar wavelet are generally similar to, and in some cases more physically reasonable than, those produced by the other wavelets (Fig. S8). These findings support the use of the Haar wavelet as a simple and effective choice for this application. We note that continuous wavelet transforms are not considered here, as they do not naturally integrate into the reconstruction-based denoising framework used in this study.

The sensitivity analysis to the other parameters can be found in Sect. S1.2. We note that in this study, the parameter selection was guided by maximizing plume detections while minimizing false positives, which may introduce the risk of overfitting to specific datasets and observing conditions. Although the chosen parameters show consistent performance across multiple scenes, their optimal values may vary under different environmental and instrumental conditions. Future work can focus on evaluating the generalizability of these parameters using larger and more diverse datasets, and on developing adaptive or region-specific tuning strategies (see Sect. ).

Furthermore, we compared the computational efficiency of the wavelet method with the DI thresholding method and a representative ML method (see more information in Sect. S1.3). Table S21 reports the total runtime and peak memory usage required to process a single scene, averaged over the same set of observations. The wavelet method does not provide an advantage over the DI method in terms of runtime or memory usage. However, because it does not require training data or a training stage, the wavelet method is still substantially more efficient than the ML-based approach.

We note that the reported values reflect only the cost of applying the optimized methods and do not include parameter tuning. In the wavelet method, the computational cost of parameter tuning scales approximately linearly with the number of parameter combinations tested.