Sensitivity of Airborne Methane Retrieval Algorithms (MF, ACRWL1MF, and DOAS) to Surface Albedo and Types: Hyperspectral Simulation Assessment

Abstract

Methane (CH₄) emissions are a major contributor to greenhouse gases and pose significant challenges to global climate mitigation efforts. The accurate determination of CH₄ concentrations via remote sensing is crucial for emission monitoring but remains impeded by surface spectral heterogeneity—notably albedo variations and land cover diversity. This study systematically assessed the sensitivity of three mainstream algorithms, namely, matched filter (MF), albedo-corrected reweighted-L1-matched filter (ACRWL1MF), and differential optical absorption spectroscopy (DOAS), to surface type, albedo, and emission rate through high-fidelity simulation experiments, and proposed a dynamic regularized adaptive matched filter (DRAMF) algorithm. The experiments simulated airborne hyperspectral imagery from the Airborne Visible/InfraRed Imaging Spectrometer-Next Generation (AVIRIS-NG) with known CH₄ concentrations over diverse surfaces (including vegetation, soil, and water) and controlled variations in albedo through the large-eddy simulation (LES) mode of the Weather Research and Forecasting (WRF) model and the MODTRAN radiative transfer model. The results show the following: (1) MF and DOAS have higher true positive rates (TP > 90%) in high-reflectivity scenarios, but the problem of false positives is prominent (TN < 52%); ACRWL1MF significantly improves the true negative rate (TN = 95.9%) through albedo correction but lacks the ability to detect low concentrations of CH₄ (TP = 63.8%). (2) All algorithms perform better at high emission rates (1000 kg/h) than at low emission rates (500 kg/h), but ACRWL1MF performs more robustly in low-albedo scenarios. (3) The proposed DRAMF algorithm improves the F1 score (0.129) by about 180% compared to the MF and DOAS algorithms and improves TP value (81.4%) by about 128% compared to the ACRWL1MF algorithm through dynamic background updates and an iterative reweighting mechanism. In practical applications, the DRAMF algorithm can also effectively monitor plumes. This research indicates that algorithms should be selected considering the specific application scenario and provides a direction for technical improvements (e.g., deep learning model) for monitoring gas emission.

1. Introduction

Atmospheric methane (CH₄) is a major greenhouse gas, contributing over 20% to current global warming. With an 84-fold higher global warming potential over a 20-year timeframe compared to CO₂, CH₄ has been increasing at an accelerating rate since 2007 [1]. Accurate quantification of atmospheric CH₄ emissions across different regions is crucial for mitigating global warming [2].

Ground-based monitoring networks, while providing precise local data, are inherently sparse and incapable of delivering the synoptic, continuous coverage required for understanding regional and global atmospheric processes [3]. Remote sensing has emerged as an indispensable tool for quantifying atmospheric CH₄ emissions across regional to global scales [4]. It enables the monitoring of diverse sources—from anthropogenic activities (e.g., oil/gas infrastructure and agriculture) to natural fluxes (e.g., wetlands and permafrost thaw)—facilitating policy-relevant source attribution [5,6,7,8]. Over recent decades, a suite of sophisticated instruments has been deployed, such as the Ozone Monitoring Instrument (OMI), the TROPOspheric Monitoring Instrument (TROPOMI), the Orbiting Carbon Observatory (OCO-2/3), and the upcoming Geostationary Environment Monitoring Spectrometer (GEMS) and Sentinel-4 missions [9,10].

Modern spaceborne spectrometers capable of hyperspectral imaging have hundreds of visible–shortwave infrared spectral channels, with spatial resolutions of tens of meters and spectral resolutions of about 10 nm [11,12,13]. They have been recognized as valuable novel instruments for detecting and measuring point-source CH₄ emissions because of their spectral and spatial resolutions. These sensors mostly function in the visible (VIS), shortwave infrared (SWIR), and ultraviolet (UV) spectral ranges, where trace gases exhibit characteristic electronic and vibrational–rotational absorption features. They gather information in the form of hyperspectral imagery, which consists of three-dimensional data cubes with spectral information from hundreds of contiguous, narrow-wavelength channels in the third dimension, combined with two sets of spatial information (across-track and along-track) in the second dimension [14,15,16,17]. This rich spectral data encodes the absorption signatures of atmospheric constituents superimposed on the solar radiation scattered by the Earth’s surface and atmosphere. The Italian Space Agency has been developing and operating PRecursore IperSpettrale della Missione Applicativa (PRISMA) since 2019. It is the first hyperspectral mission whose satellite imagery has been made publicly available to the scientific community [4,18]. Compared to a satellite, the radiance observed by an airborne platform is less affected by air molecules, aerosols, clouds, and the Earth’s surface. In recent years, advances have been made in CH₄ detection based on airborne flights. Because the aircraft AVIRIS-NG data has far higher spatial and spectral resolutions than PRISMA, CH₄ detection results are more likely to have fewer errors [13,19,20].

For CH₄ retrieval based on hyperspectral imagery, an array of algorithms has emerged, each employing distinct principles to address the inherent challenges [9,13,21,22,23]. Current operational algorithms for CH₄ retrievals, including matched filter (MF), albedo-corrected reweighted-L1-matched filter (ACRWL1MF), and differential optical absorption spectroscopy (DOAS), adopt distinct approaches to distinguish atmospheric absorption from surface contributions. For example, the MF algorithm leverages spectral covariance to enhance trace gas signals [24]; the statistical foundation assumes surface albedo spectral smoothness—an idealization violated by sharp transitions (e.g., soil–water boundaries) that introduce spurious covariance. DOAS relies on differential absorption cross-sections in specific spectral windows [25]; its accuracy degrades under low signal-to-noise ratio (SNR) conditions (typically SNR < 20 for the 2122–2485 nm CH₄ retrieval window), typical of dark surfaces, where CH₄ absorption features become indistinguishable from instrument noise.

Despite their widespread use, current evaluations of CH₄ retrieval algorithms remain fragmented, primarily focusing on clear-sky validation against sparse ground networks (e.g., TCCON) without systematic control of surface parameters [26], and systematic evaluations of their sensitivity to surface parameters remain sparse. The accuracy of remote sensing-based CH₄ concentration retrievals is inherently limited by confounding factors in radiative transfer. In addition, previous validation studies have primarily focused on algorithm performance under clear-sky conditions using validation sites (e.g., TCCON networks), yet these lack controlled variation in surface properties [21].

Among surface parameters, surface spectral properties—specifically albedo variability and land cover heterogeneity—constitute predominant sources of uncertainty [2,21,27]. This arises because terrestrial spectral characteristics modulate photon path lengths through the atmosphere, distorting the depth of CH₄ absorption features in critical bands (e.g., 1.65 μm and 2.3 μm). Surface albedo governs photon path lengths, directly impacting the detectability of weak CH₄ absorption features. Low-albedo surfaces (e.g., water and dark soils) reduce SNRs, while high-albedo targets (e.g., snow and deserts) may amplify aerosol scattering effects [28,29]. Furthermore, land cover heterogeneity generates spectral mixing that violates the plane-parallel homogeneity assumption fundamental to most one-dimensional (1D) detection algorithms [9,30].

In addition, comprehensive, systematic, and controlled comparisons of algorithm performance using a common, well-characterized dataset remain relatively scarce, particularly concerning modern high-resolution hyperspectral imagers [19,31]. Real satellite data introduces inherent confounding factors: varying illumination/viewing geometry, unknown true aerosol state, sub-pixel cloud contamination, surface heterogeneity, and calibration uncertainties. These factors make it difficult to isolate the intrinsic performance characteristics (accuracy, precision, sensitivity, robustness) of the core detection algorithms themselves [23,32,33]. Published studies often focus on comparing final satellite products or comparing algorithms applied to different sensors, where the “ground truth” is uncertain [18,34,35]. Simulations using synthetic data offer a powerful alternative for isolating surface impacts, but existing works are often limited to homogeneous scenes.

Here, we address this knowledge gap through a controlled simulation experiment using high-fidelity hyperspectral imagery (128 × 128 pixels, 72 bands). Our objectives are (1) to generate radiance scenes covering diverse land cover classes with tunable albedo; (2) analyze the sensitivity of the MF, ACRWL1MF, and DOAS algorithms to land cover types and albedos; and (3) improve CH₄ detection and test them on real AVIRIS-NG images.

2. Materials and Methods

2.1. Overview of Simulation Framework

To isolate algorithm performance from confounding surface types and albedos, we employ synthetic hyperspectral imagery as the gold standard for intercomparison. This study’s synthetic imagery covers eight distinct surface types that are modeled at two surface albedos. The framework can be seen in Figure 1, and details of the simulations can be found in Section 2.2, Section 2.3 and Section 2.4.

2.2. Reflectance Spectra at Various Albedos

The spectra chosen for this study are representative of surface types found in both natural and urban settings. These include surfaces that have historically been challenging for CH₄ mapping and common cover types observed by AVIRIS-NG. Roberts et al. [36] assembled the spectral library from which the majority of the spectra utilized to create the simulated image came. Spectra from both urban and natural environments in Southern California are included in this spectral library [19,21]. Initial spectral collection utilized AVIRIS-Classic data and Analytical Spectral Device measurements [37,38]. To align with the full-width half maxima and band centers of AVIRIS-NG, this spectral library was convolved and resampled. The specific spectra used in each class (Figure 2) are listed in

Table 1. Spectra and surface types used to generate the artificial image.

Land Cover Class	Surface Name
Confusers	Green sports field (AH)
	White painted commercial roof (WPCR)
Water	Lake
Non-Photosynthetic Vegetation (NPV)	Needle litter (NL)
Roof	Red tile roof (RTR)
Rock	Rock
Soil	bare_soil (BSoil)
Paved Surfaces	Airport asphalt (AS)

.

Then, we adjusted multispectral reflectance data to achieve the specified target albedo values (0.1, 0.4). The core principle is to adjust the albedo by globally scaling the reflectance and using weighted integration of the solar spectrum. The albedo is defined as the weighted average of the full-band solar radiation reflectance, with a weight for solar spectral irradiance:

$$Albedo={\displaystyle \frac{{\int }_{{\lambda }_{min}}^{{\lambda }_{max}}R\left(\lambda \right)E\left(\lambda \right)d\lambda }{{\int }_{{\lambda }_{min}}^{{\lambda }_{max}}E\left(\lambda \right)d\lambda }}$$

(1)

where $R\left(\lambda \right)$ is the reflectance at band $\lambda$, $E\left(\lambda \right)$ is the solar irradiance at band $\lambda$, and the integral range is between 380 and 2500 nm (visible light/near-infrared). To achieve the target reflectivity ($R_{target}\left(\lambda \right)$), we perform global linear scaling on the original reflectance ($R_{orig}\left(\lambda \right)$).

$$R_{target}(\lambda )=R_{orig}(\lambda )\times {\displaystyle \frac{{Albedo}_{target}}{{Albedo}_{orig}}}$$

(2)

2.3. Radiation Spectra Generated by MODTRAN

The reflected solar radiation spectra were simulated using the radiative transfer model MODTRAN 5.4 [39]. In this study, 1.80 parts per million (ppm) was used as the background CH₄ concentration. The background methane concentration of 1.8 ppm was selected based on MERRA-2 reanalysis data, which reports a monthly average of 1.78–1.82 ppm for the simulated region (Northern China) during 2018–2022 [40]. Assuming a sensor height of 5 km and a nadir view zenith angle, the two-way transmittance (upwelling and downwelling) for the atmospheric layers underneath the sensor was calculated. The surface reflectance option was set to 1, and the reflectance at different albedos was read, which was generated as described in Section 2.2. Thus, we can obtain the reflected radiance with various albedos and surface types (

Table 2. Simulation inputs for MODTRAN. The values are the modeled values, while the attributes are the environmental parameters influencing the radiative transfer model. With the exception of reflectance at different albedos, all characteristics remain the same.

Attribute	Values
Atmosphere MODEL	7
Atmospheric Path Type	1
Sensor height	5 km above sea level
Wavelengths	380–2500 nm
Carbon dioxide	400 ppm
Methane	1.8 ppm
Observer Zenith Angle (Deg)	180 degrees
Target Zenith Angle (Deg)	0 degrees
Observer Azimuth Angle (Deg)	0 degrees
Surface reflectance option	1 (read spectral reflectance data file)

).

The AVIRIS-NG sensor response function was then used to convolve the reflected radiation to the AVIRIS-NG wavelengths, resulting in 432 bands with a full-width half maximum (FWHM) of about 5 nm. The convolutional process based on the Gaussian function is defined as follows:

$$g\left(\lambda \right)={\displaystyle \frac{1}{\sigma \sqrt{2\pi }}}\mathrm{e}\mathrm{x}\mathrm{p}(-{\displaystyle \frac{{(\lambda -\mu )}^2}{2{\sigma }^2}})$$

(3)

where $\mu$ is the center wavelength, and $\sigma$ is the standard deviation, which determines the width of the Gaussian curve. The relationship between FWHM and $\sigma$ can be expressed as:

$$\sigma ={\displaystyle \frac{FWHM}{2\sqrt{2ln2}}}$$

(4)

2.4. Simulation of CH₄ Plume and Generation of Synthetic Datasets

Using the Weather and Research Forecasting Model large-eddy simulation (WRF-LES) at a 30 × 30 m horizontal resolution with 128 × 128 grids, we modeled an ensemble of instantaneous CH₄ column plumes. A sensible heat flux H = 100 W/m² applied uniformly to the surface was used to generate buoyant turbulence. Surface drag with a height of 0.1 m for aerodynamic roughness created mechanical turbulence. For a given emission rate and weather conditions, WRF-LES can simulate realistic three-dimensional volume mixing ratio enhancements [7,41]. We can produce artificial 2D enhanced images from the 3D LES fields at each time step that resemble real observations from a particular remote sensing device. Integrating over the column yields two-dimensional CH₄ plumes (Figure 3).

We constructed artificial plumes from sources with flux rates spanning orders of magnitude, assuming that any given 2D image of enhancements at a reference flux rate may be scaled by an arbitrary factor to reflect different emission rates. The total mass of CH₄ in the plume is scaled to obtain the emission rate [42]. In this study, we concentrated on plume emission rates between 500 and 1000 kg/h, which is the region where most common CH₄ point sources were detected [32]. We focused on 500–1000 kg/h because this range represents mid-intensity industrial sources (e.g., small oil/gas well pads, landfill vents) that are widespread but often under-monitored [32,43]. While GHGsat (and similar satellites) can detect sources as low as ~100 kg/h [31], these ultra-low-emission sources typically require pixel-level SNR > 30, which is challenging for airborne sensors like AVIRIS-NG in low-albedo scenarios (e.g., water, dark soil). We mainly concentrated on performing LES experiments with different concentration magnitudes, ignoring wind speed, even though it is the primary factor influencing distinct spatial patterns of the total CH₄ column and the relationship between the total CH₄ concentration within a distinct plume and the associated CH₄ flux rate. With the wind speed maintained at 3 m/s, our training data will be an image with the augmentation of the entire column in every pixel. Our LES trials produced 240 simulated plumes in total.

Furthermore, the two-way absorption must be considered when calculating the unit CH₄ absorption coefficient. By multiplying the HITRAN absorption cross-sections (${\sigma }_H$; [44]) by the CH₄ volume mixing ratio enhancement (∆VMR) and vertical column density of dry air (VCD) below 5 km from the MERRA-2 meteorological reanalysis [40], we calculate the optical depth of the CH₄ plume τ(λ) at wavelength λ:

$$\tau \left(\lambda \right)={\sum }_i^{72}∆{\mathrm{V}\mathrm{M}\mathrm{R}}_i{\mathrm{V}\mathrm{C}\mathrm{D}}_i{\sigma }_H,i(\lambda )$$

(5)

where the number 72 corresponds to the vertical grid resolution of the WRF-LES model, which divides the 5 km atmospheric column (below the sensor) into 72 evenly spaced layers to resolve vertical variations in methane volume mixing ratio (∆VMR) and dry air vertical column density (VCD). In addition, MERRA-2 data were selected over CAMS due to its finer vertical resolution (72 layers) matching WRF-LES, hourly temporal alignment, and validated accuracy (<3% bias) for the simulated region [45]. CAMS was not used due to coarser vertical resolution (>100 m near the surface) and 3-hourly temporal intervals, which would introduce interpolation errors. According to Beer’s law, the backscattered solar radiation’s plume transmission $T\left(\lambda \right)$ is equal to the negative exponential of $\tau \left(\lambda \right)$ weighted by the geometric air mass factor (AMF) A (Figure 4):

$$T\left(\lambda \right)=\mathrm{e}\mathrm{x}\mathrm{p}\{-A\tau \left(\lambda \right)\}$$

(6)

Each pixel’s radiance spectrum is multiplied by this additional plume transmission. In addition, we added noise to the radiance based on the SNR [46]. Finally, we obtained the synthetic datasets (

Table 3. The specifications of synthetic radiance.

Parameter	Range
Surface albedo	0.1, 0.40
Land cover types	Confusers/Water/Rock/Non-Photosynthetic Vegetation/Paved Surfaces/Roof/Soil
Emission rates	500\1000

; Figure 5).

2.5. Methods for Detection of CH₄

The highest CH₄ absorptions occur at approximately 1650 and 2300 nm. The features of CH₄ absorption are similar to those of the other two interfering gases. The CH₄ absorption characteristic at about 2300 nm is frequently selected as the retrieval window because it is less vulnerable to the influence of other gases. Reflected light within the CH₄ retrieval window (2122–2485 nm) was examined in this work.

2.5.1. Matched Filter

In earlier spaceborne detection studies, the CH₄ column enhancement (∆XCH4) was frequently recovered using the MF approach. CH₄ point sources are frequently retrieved using this technique.

Using the spectral signature of CH₄ absorption s (s is the unit CH₄ absorption coefficient, which is extracted from radiance simulated with the MODTRAN model, as previously described [24,47], the sensor’s measured radiance is modeled as a function of the concentration of CH₄ α (α is the CH₄ column enhancement) and a target spectrum t. The background is corrupted by zero-mean additive colored Gaussian noise with covariance C. Let $L_0$ be the ambient at-sensor radiance with background concentrations of an absorbing gas but no enhancement. CH₄ enhancement’s impact is determined using the Beer–Lambert absorption law, $L(\alpha , s) = L_0{e}^{-\alpha s}$. A first-order Taylor series expansion is used to linearize the Beer–Lambert absorption equation, resulting in a quadratic optimization problem:

$$L_0{e}^{-\alpha s}=L_0-\alpha t_s\left(L_0\right)$$

(7)

It is possible to generate the radiance-dependent target spectrum $t_s\left(L_0\right)$ using radiative transfer simulations of transmittance, radiative transfer simulations of changes in radiance with changes in CH₄ concentration, or the absorption spectrum of CH₄ [48].

Since $L_0$ is the unknown nonenhanced ambient at-sensor brightness, we estimate it using the mean at-sensor radiance $\mu$.

Since we are interested in a single gas with constant characteristic absorption $s$, we eliminate the subscript in $-t_s$ (simply represented by t) and assume negative values in order to simplify the notation. This leaves the target spectrum as simply t. The aim is to reduce the residuals between the modeled and observed spectra. The following is an expression for its least-squares solution:

$${\widehat{\alpha }}_i={\displaystyle \frac{{\left(L_i-\mu \right)}^T{C}^{-1}(t(\mu ))}{{\left(t\left(\mu \right)\right)}^T{C}^{-1}(t(\mu ))}}$$

(8)

It is important to remember that the linearization in Equation (7) only works when the CH₄ plume is weak and has a small α [43,49]. For a large α, the retrieval result will be inaccurate. The following quantification of emission rates is impacted by this linearization, even though it has no effect on the detection of CH₄ point sources.

2.5.2. The Albedo-Corrected Reweighted-L1-Matched Filter

A new computationally efficient technique is presented by Foote et al. [24] that matches the filter retrieval of the trace gas concentration path length by using sparsity and an albedo correction. Every reweighting iteration in the reweighted-L1 approach necessitates a solution to the L1-regularized problem. Furthermore, it is more challenging to identify trace gas enhancement in images taken over low-albedo surfaces. The absorption signal in terms of absolute radiance drops as surface albedo decreases, while trace gas absorption—as a percentage of ambient at-sensor light—remains independent of albedo. The target spectrum, which incorporates surface reflectance and represents the change in brightness for a unit change in CH₄ enhancement, is theoretically constructed to handle this albedo impact. It might be more accurate to estimate the target spectrum per pixel while taking local albedo into consideration. Because the absorbed radiance in the Beer–Lambert transmission is directly reliant on the initial brightness in the absence of the absorber, Foote et al. [24] adjust the target spectrum by a pixel’s observed albedo factor to compensate for this non-specificity effect in our optimization. The spectral mean $\mu$ and the radiance spectrum $L_i$ of the ith pixel are used to compute the scalar albedo factor $r_i$:

$$r_i={\displaystyle \frac{L_i^T\mu }{{\mu }^T\mu }}$$

(9)

The approach includes a process for iteratively determining the best gas enhancement solution for this albedo-corrected solution, the sparse solution, and the original MF. The ACRWL1MF technique improves CH₄ enhancement retrieval on low-albedo surfaces.

2.5.3. Differential Optical Absorption Spectroscopy

The purpose of DOAS is to distinguish between low-frequency information (slowly changing part) caused by Rayleigh scattering and Mie scattering and high-frequency information (fast-changing part) caused by optical thickness changes induced by the rapidly varying molecular absorption characteristics of the wavelength. Numerous atmospheric components have been recovered using this technique, and the absorption cross-section can be written as follows:

$${\sigma }_j\left(\lambda \right)={\sigma }_{j_o}\left(\lambda \right)+{\sigma }_j^{\mathrm{\text{'}}}(\lambda )$$

(10)

where ${\sigma }_j^{\mathrm{\text{'}}}(\lambda )$ is the fast-rotating portion, mostly originating from different absorbing gas molecules, and ${\sigma }_{j_o}$ is the absorption cross-section derived from low-frequency information, which is predominantly due to Rayleigh and Mie scattering. Methane’s absorption cross-section (used in DOAS) is strongly dependent on temperature and pressure: lower temperatures narrow absorption lines, while higher pressures broaden them via collisional effects [25,50]. Neglecting these effects can introduce up to 15% bias in retrieved concentrations [25]. Radiation intensity in the atmosphere is influenced by three primary mechanisms—Rayleigh scattering, Mie scattering, and the fast absorption of trace gases:

$$I\left(\lambda \right)=I_0(\lambda )·\mathrm{e}\mathrm{x}\mathrm{p}[-L({\sum }_j\left({\sigma }_j^{\mathrm{\text{'}}}\left(\lambda \right)·c_j\right))]·\mathrm{e}\mathrm{x}\mathrm{p}[-L({\sum }_j\left({\sigma }_{j_o}\left(\lambda \right)·c_j\right)+{\epsilon }_R\left(\lambda \right)+{\epsilon }_M\left(\lambda \right))]·A(\lambda )$$

(11)

where $c_j$ is the concentration of the jth gas, $A\left(\lambda \right)$ is the influence of systemic and atmospheric turbulence, and ${\epsilon }_R\left(\lambda \right)$ and ${\epsilon }_M\left(\lambda \right)$ are the extinction effects of Rayleigh and Mie scattering, respectively. The radiation change $I^{\mathrm{\text{'}}}\left(\lambda \right)$ brought on by low-frequency information can be approximated by a polynomial in this algorithm to make computations easier:

$$I^{\mathrm{\text{'}}}\left(\lambda \right)=I_0(\lambda )·\mathrm{e}\mathrm{x}\mathrm{p}[-L({\sum }_j\left({\sigma }_{j_o}\left(\lambda \right)·c_j\right)+{\epsilon }_R\left(\lambda \right)+{\epsilon }_M\left(\lambda \right))]·A(\lambda )$$

(12)

This is one of the most popular techniques for eliminating the impact of low-frequency data. Next, we can express the differential optical thickness D′ as follows:

$$D^{\mathrm{\text{'}}}=ln{\displaystyle \frac{I_0^{\mathrm{\text{'}}}\left(\lambda \right)}{I\left(\lambda \right)}}=L({\sum }_j({\sigma }_{j_o}\left(\lambda \right)·c_j))$$

(13)

The gas concentration can be obtained by combining many bands and using differential absorption spectra to match the gas absorption cross-section using least squares.

2.6. Statistical Analysis

Performance indicators obtained from the confusion matrix were used to quantitatively assess the correctness of the inversion findings. In particular, the following metrics were computed: true positives (TP), which are instances accurately recognized as the target class; true negatives (TN), which are instances accurately classified as not belonging to the target class; false positives (FP), which are instances mistakenly classified as belonging to the target class; and false negatives (FN), which are instances mistakenly classified as not belonging to the target class.

These fundamental counts were used to compute the F1 score, a robust metric for assessing classification performance, particularly when class distributions may be imbalanced. The F1 score is the harmonic mean of recall and precision and is defined as follows:

$$\mathrm{F}1=2{\displaystyle \frac{Precision\times Recall}{(Precison+Recall)}}$$

(14)

where Precision (Positive Predictive Value) is equal to TP/(TP + FP), which measures the proportion of identified positives that are actual positives; Recall (sensitivity, or true positive rate) is equal to TP/(TP + FN), which measures the proportion of actual positives that are correctly identified.

3. Results

3.1. Subsection

Sensitivity of Retrieval Algorithms to Surface Types: Plume Detection Capabilities

MF, ACRWL1MF, and DOAS are actually retrieval algorithms (designed to quantify CH₄ concentrations), but this section focuses on their ability to detect plumes (i.e., identify pixels with CH₄ enhancements), using detection (F1) metrics. Figure 6 presents a comparative analysis of gas plume detection performance using three distinct methods, MF, ACRWL1MF, and DOAS, against a sample of the simulated plume (at 03:30:30, with emission rate: 500 kg/h; albedo: 0.4; land surface type: AH). We can see that the retrieval results obtained using the MF and DOAS methods achieve higher values for TP (93.5% for MF; 92.0% for DOAS), but lower values for TN (51.55% for MF; 49.9% for DOAS) and F1 (0.072 for MF; 0.069 for DOAS). The results obtained using the ACRWL1MF method achieve higher values for TN (95.9%) and F1 (0.347), but lower values for TP (63.8%). In addition, the CH₄ column concentrations retrieved by MF tend to be higher than the actual values (Figure 6b).

We further quantitatively evaluate the accuracy of the retrieved concentration distribution (Figure 7). We can see that the results retrieved using ACRWL1MF are relatively consistent with the reference plume. The results retrieved using MF and DOAS have lower values in their distributions.

Figure 8 systematically evaluates the robustness of the three retrieval algorithms—MF, ACRWL1MF, and DOAS—across diverse vegetation types under varying emission rates (500/1000 kg/h) and surface albedos (0.1/0.4). MF and DOAS remain stable despite albedo/emission changes, with minor F1 fluctuations (<5%) between panels (a–d), but they obtain lower F1 scores across all vegetation types (e.g., needle litter, bare soil, and green sports field). However, although ACRWL1MF has a higher F1value, it exhibits a relatively high vegetation dependence especially at 500 kg/h and the F1 value is higher at 1000 kg/h than at 500 kg/h. Therefore, it indicates that different algorithms vary in their vegetation dependence.

3.2. Sensitivity of Algorithms to Surface Albedo

We tested the sensitivity of algorithms to surface albedos (0.1 vs. 0.4) by comparing the retrieval F1 scores of the three algorithms—MF, ACRWL1MF, and DOAS at a certain albedo (Figure 9). We found differences between the retrieved plume concentrations at different albedos, especially for the MF method (p = 0.003) and ACRWL1MF method (p = 0.01). ACRWL1MF-retrieved CH₄ at an albedo of 0.1 is a little higher than that at an albedo of 0.4 (Figure 9b), indicating that ACRWL1MF can achieve better results at low albedos. However, DOAS are less sensitive to surface albedos (p > 0.05).

3.3. Sensitivity of Algorithms to Different Emission Rates

Furthermore, we analyzed the sensitivity of algorithms to different emission rates (high emission rates generally mean higher plume concentrations). From Figure 10, we can see that the three algorithms have better F1 values at a higher emission rate (1000 kg/h). The F1 scores for MF and DOAS are consistently below 0.10 at all emission rates, with <10% variation between 1000 kg/h (dark bars) and 500 kg/h (light bars).

ACRWL1MF has a significant reduction in F1 (from about 0.8 to about 0.4) from high emission rate (1000 kg/h) to a lower emission rate (500 kg/h). Compared to the results at albedo of 0.2, MF-retrieved results at albedo of 0.4 have a significant difference between emission rates of 500 kg/h and 1000 kg/h (p = 0.00025). However, DOAS-retrieved results have no significant difference between emission rates of 500 kg/h and 1000 kg/h (p > 0.05). These results demonstrate that the three algorithms’ emission rate sensitivity varied: Both surface albedo and emission rates affected the ACRWL1MF method’s ability to detect plumes, whereas the DOAS method was not affected by either of these factors. At higher emission rates, surface albedo primarily affected the MF method’s ability to detect plumes.

In addition, we analyzed the algorithm performance’s sensitivity to the proportion of actual plume pixels (Panel A) under fixed conditions (03:30:30, 500 kg/h, albedo 0.4, AH surface). Actual plume coverage varies widely (0.5–12.5% across samples). The F1 curves for MF and DOAS have a similar trend to the proportion of plume-containing pixels. The scatterplot reveals a strong linear correlation between the MF-retrieved (or DOAS-retrieved) plume and the proportion of actual plume pixels to all pixels (R² > 0.95). However, for the ACRWL1MF-retrieved results, F1 is not related to the proportion of actual plume pixels to all pixels (R² < 0.1). This indicates that the ACRWL1MF retrieval method is less sensitive to plume pixels (Figure 11 and Figure 12).

In addition, we tested the algorithm performance sensitivity to the proportion of actual plume pixels (Panel A) for all plumes. We can see that the DOAS-retrieved method has a low and unstable R² value at 1000 kg/h compared to that at 500 kg/h, whereas the MF and ACRWL1MF retrieval methods have relatively stable R² values (about 0.95 for MF, about 0.1 for ACRWL1MF).

3.4. Improvement in CH₄ Retrievals and Application to Real AVIRIS-NG Image

DOAS requires the careful selection of the fitting window and reference spectrum, and converting SCD to VCD requires an additional AMF, introducing another potential error source. Additionally, the above three algorithms have different advantages and disadvantages and are suitable for different scenarios. Given these observations, we developed a new method named DRAMF (Dynamically Regularized Adaptive Matched Filter, Algorithm 1) by combining the MF and ACRWL1MF methods. DRAMF uses dynamic background updates and reweighted regularization (the core principle of ACRWL1MF) during the iteration process. The pseudocode is as follows:

Algorithm 1. DRAMF

The DRAMF has been validated in simulation tests. We also present a comparative analysis of gas plume retrieval performance using DRAMF against a sample of the simulated plume (at 03:30:30 with emission rate: 500 kg/h; albedo: 0.4; land surface type: AH). From Figure 13, we can see that the algorithm successfully detected the plume concentration with a higher F1 value (0.129), a higher TP value (81.4%), and a higher TN value (78.2%), with a 180% increase in F1 compared to MF and DOAS and a 128% increase in TP compared to ACRWL1MF (Figure 6).

In addition, we apply DRAMF method to the real AVIRIS-NG image (subsets of flight lines ang20160211t075004 and ang20170906t210217). From Figure S1, we can see that DRAMF method can effectively monitor plumes, and the plumes around industrial facilities identified by the DRAMF method (Figure 14c) are consistent with the “diffused plume” characteristics reported in the literature (Figure 14, [24,51]). In addition, to confirm physical rationality of plume, we calculated the wind direction and speed from the ERA5 reanalysis dataset, obtained from the Copernicus Climate Data Store (CDS). The 10 m u-component of wind and 10 m v-component of wind were used for this derivation.

Figure 14c exhibits a diffused plume morphology, characterized by a downstream spread of approximately 1000 m and a gradual concentration decline from ~300 ppb near the source (an industrial facility) to ~50 ppb at the edge. This pattern is consistent with Foote et al. [24] (Figure 14b). Furthermore, the smooth concentration gradient without abrupt jumps supports the validity of the plume, as such discontinuities are typical artifacts of surface reflection, a finding in agreement with Thorpe et al. [13]. Figure 14f identifies the plume origin at a small industrial vent and its downstream extension of approximately 150 m. This spatial scale is consistent with the characteristics of a “low-intensity diffused plume” as described in Foote et al. [51].

However, the locally high concentration signals in the rooftop areas of Figure 14c,f marked by yellow ellipses may be influenced by surface reflection artifacts. The localized high concentrations in Figure 14c,f are likely surface artifacts. While industrial rooftop ventilation systems are common emission sources in the region, analysis of the concentration time series from adjacent pixels showed no spatial gradient—consistent with random surface artifacts rather than a true emission plume. Although the rooftop signal in Figure 14c,f may be an artifact, not a real emission, and does not conflict with the literature conclusions. Subsequent analysis will use high-resolution land cover data (e.g., high-resolution imagery from Google Earth) to eliminate misinterpretations in non-plume areas, and quantitative analysis will be conducted only for areas with clearly defined plume patterns.

On the simulated data surface, although the new algorithm achieves good F1, TP, and TN results under different conditions, it may also have uncertainties in accurate quantification of plume concentration. Figure S1 systematically evaluates the robustness of DRAMF—across diverse vegetation types at varying emission rates (500/1000 kg/h) and surface albedos (0.1/0.4). Compared to MF and DOAS (Figure 8), DRAMF generally exhibits a higher F1 value. Similarly, DRAMF exhibits a higher F1 value at higher emission rates. However, DRAMF method showed a relatively large vegetation dependence, which can be seen from the fluctuations among Figure panels (a) and (d).

Furthermore, we tested the algorithm performance sensitivity to the proportion of actual plume pixels for all plumes (Figure S1). We can find that the DOAS retrieval method has a low and unstable R² value at 1000 kg/h compared to that at 500 kg/h. It demonstrated that DRAMF is still affected by the proportion of plumes in images with a low emission rate (corresponding to low concentration). The results also showed that the AH and WPCR’s F1 values were generally different from other surface types (Figure 14) and the AH and WPCR’s R² values of the relationship between F1 of the different algorithms and the ratio of actual plume pixels were unstable, suggesting that the confusers’ spectra, comprising hydrocarbons or comparable SWIR absorption, might have obstructed the CH₄ absorption characteristics. Furthermore, DRAMF also marginally improves with higher emissions (Figure S2). In total, the above results indicate that although DRAMF has a better performance, it also has some limitations in different scenes.

4. Discussion

Numerous algorithms have been developed to tackle the retrieval problem, and each has distinct theoretical foundations, strengths, limitations, and computational demands [22,24,26,30]. Extracting quantitative trace gas concentrations (e.g., VCDs) from measured top-of-atmosphere or top-of-canopy radiance spectra is a formidable inverse problem. Distinguishing the relatively weak, spectrally structured absorption signal of a specific trace gas from this complex background requires sophisticated physical models and robust mathematical retrieval algorithms [5,32]. The accuracy and precision of these retrievals directly impact the scientific validity and utility of the derived data products for downstream applications.

This study systematically evaluated the sensitivity of three mainstream CH₄ detection algorithms (MF, ACRWL1MF, and DOAS) to surface type, albedo, and emission rate through controlled simulation experiments, and proposed an improved DRAMF algorithm. For detection performance (plume identification), MF and DOAS have higher true positive rates (TP > 90%) in high-reflectivity scenarios, but the problem of false positives is prominent (TN < 52%); ACRWL1MF significantly improves the true negative rate (TN = 95.9%) through albedo correction but lacks the ability to detect low concentrations of CH₄ (TP = 63.8%); DRAMF achieved an F1 score of 0.129—180% higher than MF (0.072) and DOAS (0.069)—and a TP value of 81.4%, which is 32% higher than ACRWL1MF (63.8%). The research results not only reveal the performance characteristics and limitations of existing algorithms but also provide a theoretical basis and technical path for high-precision CH₄ remote sensing inversion under complex surface conditions.

4.1. Different Sensitivities of Retrieved Algorithm to Surface Parameters

The evolution of DOAS has spanned decades, from ground-based MAX-DOAS to satellite sensors such as GOME, SCIAMACHY, OMI, and TROPOMI [13]. DOAS excels in computational efficiency and robustness for strong absorbers with clear differential features but faces challenges with weak absorbers, strong aerosol interference, and complex surface albedo, and it requires careful selection of the fitting window and reference spectrum [21,30]. Converting SCD to VCD requires an additional AMF calculation using a radiative transfer model (RTM), introducing another potential error source. MF’s strengths lie in handling complex scenes (high aerosol load, varying albedo) and providing multiple parameters simultaneously [51]. However, it suffers from extremely high computational intensity (requiring massive pre-computed databases), potential “database representation errors” if the true state is not well-sampled, and difficulties in characterizing retrieval errors and sensitivities due to its discrete nature. For AVIRIS-NG data (432 bands, 128 × 128 pixels per scene), a database covering 8 surface types, 2 albedo levels (0.1, 0.4), and 2 emission rates (500–1000 kg/h) requires ~300 GB of storage (uncompressed). Generating this database takes ~6–10 h on a NVIDIA A100 GPU, as it involves MODTRAN radiative transfer simulations for each parameter combination [39]. This is a practical limitation for real-time applications (e.g., airborne in-flight processing). In addition, MF’s application has been largely tied to specific sensors, such as OMI. ACRWL1MF’s strengths include rigorous error propagation, simultaneous retrieval, which reduces cross-talk, flexibility in state vector definition, and applicability to various sensors [24]. Its main challenges are high computational cost per retrieval and sensitivity to the choice of a priori constraints and forward model accuracy.

All three traditional algorithms (MF, ACRWL1MF, DOAS) rely on the WRF-LES model to simulate plume dispersion, as none explicitly model wind/turbulence in their retrieval logic. WRF-LES accounted for atmospheric mixing via two mechanisms: (1) buoyant turbulence from a uniform sensible heat flux (H = 100 W/m²) and (2) mechanical turbulence from surface drag (aerodynamic roughness height = 0.1 m) [52]. The fixed wind speed (3 m/s) was chosen to isolate the impact of surface albedo/type on algorithm performance, rather than confounding variables like wind-driven plume shape changes. Since all algorithms used the same WRF-LES-generated plume fields, differences in performance (e.g., ACRWL1MF’s higher TN vs. MF’s higher TP) are attributed to their spectral processing logic, not variations in dispersion modeling. While the fixed wind speed (3 m/s) ensured controlled testing of surface effects, wind variability directly impacts α (CH₄ column enhancement) by altering plume dispersion: higher wind speeds (e.g., 8 m/s) spread plumes horizontally, reducing α (lower column concentration) in individual pixels, while lower wind speeds (e.g., 1 m/s) increase α (higher local concentration). This affects MF’s linearization (Equation (7)), as the approximation $L_0{e}^{-\alpha s}=L_0-\alpha t_s\left(L_0\right)$ becomes less accurate for large α (low wind) or small α (high wind). Future work will evaluate algorithm performance across wind speeds (1–8 m/s) to quantify this sensitivity.

The research results clearly indicate that different algorithms have significant differences in their sensitivity to surface types and albedo. MF and DOAS remain stable despite albedo/emission changes, but they obtain lower F1 scores across all vegetation types (e.g., needle litter, bare soil, and green sports field). However, although ACRWL1MF has a higher F1value, it exhibits a relatively high vegetation dependence especially at 500 kg/h (Figure 8). In addition, MF and DOAS exhibit high true positive rates (TP > 90%) in high-albedo scenarios but are accompanied by high false positive rates (TN < 52%), which is consistent with the false covariance problem caused by the failure of the “smooth surface hypothesis” pointed out in previous studies [21,30]. In contrast, ACRWL1MF significantly improves the true negative rate (TN = 95.9%) by introducing albedo correction and reweighting mechanisms, but at the cost of a decrease in the true positive rate (TP = 63.8%; Figure 6). This result validates the advantage of ACRWL1MF in handling complex surface reflectance variations [9,11].

The impact of surface albedo on algorithm performance exhibits significant algorithm dependence. MF’ and ACRWL1MF retrieved results have significant differences between albedo of 0.1 and albedo of 0.4, while DOAS are relatively insensitive to changes in albedo. In addition, ACRWL1MF-retrieved CH₄ at an albedo of 0.1 is a little higher than that at an albedo of 0.4 (Figure 9). This phenomenon can be attributed to the ACRWL1MF algorithm’s ability to enhance weak signals, and the low SNR in low-albedo scenarios is the main challenge of traditional algorithms [28,29]. It is worth noting that all these algorithms have better F1 scores at high emission rates (1000 kg/h) than at low emission rates (500 kg/h), reflecting that the current remote sensing technology still needs to improve its ability to detect weak CH₄ signals.

4.2. Algorithm Limitations and Improvement Directions

This study reveals the inherent limitations of existing algorithms in dealing with complex surface conditions. MF and DOAS have a weak dependence on the surface type, while ACRWL1MF are prone to misjudgment in scenarios with heterogeneous surfaces. Although ACRWL1MF has improved robustness, its ability to detect low concentrations of CH₄ is insufficient. These limitations mainly stem from the following aspects: (1) simplified modeling of surface reflectance: most algorithms assume that surface reflectance has spectral smoothness, which is often not the case in practical applications (changes in vegetation cover, water boundaries, etc.); (2) idealization of RTM: the one-dimensional RTM ignores the three-dimensional atmospheric structure and sub-pixel heterogeneity, resulting in inversion errors [30]; (3) signal noise interference: under conditions of low-albedo surfaces and weak emission sources, CH₄ absorption characteristics are easily overwhelmed by instrument noise, reducing the sensitivity of the algorithm.

To accomplish effective mapping of global methane concentration, some studies have created collaborative multi-source satellite data intelligent remote sensing inversion methods for atmospheric methane concentration (e.g., UNMAMO; [28,45]). Although the band ratio inversion strategy can significantly reduce the inversion errors introduced by factors like surface albedo, aerosols, and altitude [45], the efficiency and success rate of greenhouse gas concentration inversion are severely limited by the use of fully physical optimal estimation algorithms, which are based on online simulation of radiative transfer and have high computational costs and numerous algorithm limitations. In addition, the existing machine learning inversion strategy driven by both data and mechanism generally streamlines the iterative optimization process, which brings great uncertainties to the retrieved results.

In response to the above issues, the DRAMF algorithm proposed in this study has achieved performance improvements through the following innovative design aspects: (1) dynamic background update mechanism: adjusting the background model in real time to adapt to the spatiotemporal changes in surface reflectance; (2) iterative reweighting regularization: introducing adaptive weights into the matched filtering process to enhance the detection capability of weak signals; (3) multi-scale spatial constraints: combining spectral and spatial information to prevent false detections and preserve true signals. The experimental results show that DRAMF significantly outperforms traditional algorithms in F1 score (0.129) and TP value (81.4%), showing especially strong robustness in low-emission scenarios. While this study focuses on advancing filter-based algorithms (MF, ACRWL1MF, DRAMF), it is important to note that existing ML methods (e.g., CNNs for plume detection [53]; transformer models for concentration quantification [54]) excel at handling high-dimensional hyperspectral data and sub-pixel heterogeneity. DRAMF complements these ML approaches by retaining physical interpretability (via Beer–Lambert law and dynamic background modeling), which is often lacking in “black-box” ML models. For example, DRAMF’s iterative reweighting mechanism is more transparent than ML’s feature extraction, making it easier to diagnose errors (e.g., false positives from surface albedo variations). A direct head-to-head comparison is planned for future work, focusing on low-emission scenarios (<200 kg/h) where ML often struggles with data scarcity.

4.3. Actual Application Potential and Challenges

Although this study validated the algorithm’s performance through simulation experiments, applying it to actual remote sensing data still faces many challenges [32,46]. Firstly, simulated data cannot fully capture complex factors in real scenes, such as cloud pollution, aerosol uncertainty, and instrument calibration errors. Secondly, the detection capability of algorithms for small-scale leakage sources needs to be further improved, especially in areas with dense industrial facilities, where sub-pixel mixing effects may lead to missed detections [28,34,35,43]. In addition, the computational efficiency of algorithms is a key issue in practical applications, and the iterative calculation requirements of dynamic regularization processes may limit their practicality in processing large amounts of data.

In actual AVIRIS-NG data testing, DRAMF successfully detected CH₄ emission hotspots consistent with the literature [24,51], but some missed detections occurred in low-concentration areas. This suggests that future research needs to combine high-resolution satellite data (such as PRISMA or EnMAP) with advanced machine learning techniques (or deep learning approaches) to further optimize the algorithm structure and improve the ability to recognize weak CH₄ signals. In addition, to further improve performance for <200 kg/h sources, future iterations of DRAMF will integrate a “weak-signal boost” module, which weights spectral bands with the strongest CH₄ absorption (e.g., 2300 nm) more heavily, reducing noise interference.

The scientific contribution of this study lies in quantifying the impact of surface parameters on different CH₄ inversion algorithms through systematic control experiments for the first time and proposing targeted improvement plans. The research results are of great significance for guiding regional-scale CH₄ emission monitoring. Algorithm selection strategy: In high-albedo and heterogeneous surface areas, it is recommended to use ACRWL1MF or DRAMF; in low-reflectivity scenarios and fewer plume pixels, ACRWL1MF performs better, while in high-reflectivity scenarios and more plume pixels, DRAMF performs better; DOAS can maintain the stability of inversion results under different surface types, albedo, etc. Additionally, we compiled a table (Table S1) that ranked common operational requirements (such as regional monitoring and industrial leak detection) and compiled the preferred algorithms for various situations (such as albedo, emission rate, and plume size). Looking ahead, with the successive launches of hyperspectral satellite constellations such as Sentinel-4 and GEMS, as well as continuous innovation in algorithm theory, atmospheric CH₄ remote sensing monitoring will play a more important role in global climate change research and emission reduction policy formulation.

5. Conclusions

This study systematically evaluated the sensitivity of existing CH₄ inversion algorithms to surface parameters through high-fidelity simulation experiments and actual data validation, and proposed an improved DRAMF algorithm. The results show that the surface type and albedo significantly affect the accuracy of CH₄ inversion. The traditional MF approach and DOAS have high true positive rates (TP > 90%) in high-albedo scenarios, but the problem of false positives is prominent (TN < 52%). The ACRWL1MF increases the true negative rate to 95.9% by introducing surface reflectance constraints, but its ability to detect weak CH₄ signals is insufficient (TP = 63.8%). The F1 scores of all algorithms under high emission rates (1000 kg/h) increased compared to those under low emission rates (500 kg/h), indicating that the current technology still needs improvement in this aspect. In addition, the improved DRAMF algorithm significantly enhances adaptability to complex scenes through dynamic background updates and iterative reweighting regularization. DRAMF outperforms traditional algorithms in overall performance, with an F1 score of 0.129, which is 180% higher than that of MF and DOAS, and a TP value (81.4%) that is 32% higher than that of ACRWL1MF. Algorithms should be selected considering the specific application scenario. The practical application potential and limitations of AVIRIS-NG data verification show that DRAMF can effectively identify CH₄ leaks from industrial sources, but missed detections still occur in low-concentration areas. This study analyzed the impact of surface parameters on CH₄ inversion algorithms and proposed an improved algorithm that provides a technical reference as a means for monitoring regional-scale CH₄ emissions.