To identify regions influenced by CH 4 emissions, a UAV–AirCore sampling system was deployed to conduct targeted atmospheric observations over the oilfield areas. This stage focused on screening regions exhibiting significant CH 4 concentration enhancements and providing spatial observational constraints for subsequent emission analysis.

The UAV–AirCore platform consisted of a DJI Inspire Pro 1 unmanned aerial vehicle, a custom-developed active AirCore sampling unit, a GPS-based positioning and time-logging module, and a Picarro G2401-m cavity ring-down spectroscopy (CRDS) gas analyzer for post-flight concentration analysis. The AirCore system enables continuous air sampling along the UAV flight path, with a CH 4 measurement precision of approximately 1 ppb. Typical flight altitudes ranged from 20 to 200 m above ground level, mainly constrained by UAV endurance and local airspace regulations. The system provides high spatial flexibility and multi-height sampling capability, enabling detailed characterization of methane concentration distributions in complex oilfield environments. To enable rapid and accurate air sampling under practical flight conditions while minimizing the risk of gas contamination, several hardware optimizations were implemented. These included inert gas conditioning and inertial optimization of the sampling pathway to reduce adsorption effects, lightweight integration of the AirCore unit to improve flight stability, and simplification of analytical interfaces to enhance temporal synchronization between gas samples and positional records. These improvements collectively enhanced sampling efficiency and measurement reliability.

Given the limited flight duration and restricted airspace typical of oilfield environments, an optimized flight trajectory was defined as one that maximizes spatial coverage and sampling density over potential plume-affected areas within a single flight while satisfying safety and endurance constraints. To achieve this objective, a hybrid flight path planning algorithm was developed by integrating an improved Ant Colony Optimization (ACO) algorithm with the Group Spider Optimization (GSO) algorithm (Deng et al., 2019; Luque-Chang et al., 2018). A dynamic guiding factor was introduced into the ACO framework to accelerate convergence, while the GSO component enhanced global search capability and prevented premature convergence. This hybrid strategy enabled efficient identification of near-optimal flight trajectories in large and constrained search spaces.

During field operations, the UAV–AirCore system first conducted horizontal scanning flights to characterize near-surface CH 4 concentration patterns across the oilfield area. Horizontal flights were performed at altitudes of approximately 50–80 m above ground level, with flight tracks arranged along and perpendicular to the prevailing wind direction. The spacing between adjacent flight tracks was typically 30–50 m, and the flight speed was maintained at 3–5 m s − 1 to ensure stable AirCore sampling. Continuous CH 4 concentration measurements were collected and georeferenced, enabling the identification of spatially coherent concentration enhancements relative to the local background. Based on the identified anomalous regions, top-down three-dimensional spiral flights were subsequently conducted to characterize the vertical structure of methane plumes. Vertical flights generally covered altitudes from near the surface up to 150–200 m. Spiral trajectories were designed with radii of approximately 30–60 m, vertical sampling intervals of about 5–10 m, and ascent/descent rates of 0.5–1.0 m s − 1 . Each vertical flight consisted of 2–4 spiral laps and lasted approximately 10–15 min. This sampling strategy provided vertically resolved CH 4 concentration profiles, which supplied essential observational constraints for subsequent emission modeling and inversion analysis.

After identifying the spatial locations of anomalous CH 4 concentrations using the UAV–AirCore system, the Emission-Partition model was applied to reconstruct the methane concentration field and estimate the corresponding emission rates within the study area. In atmospheric methane monitoring, the concentration measured at a given sampling location is generally considered to consist of a regional background component and concentration enhancements resulting from the atmospheric dispersion of multiple emission point sources. Therefore, a multi-source superposition model is required to accurately represent the observed CH 4 concentrations.

In this study, a three-dimensional Gaussian dispersion model is employed to simulate the concentration enhancement produced by each individual emission point source. Under the assumption of linear superposition, the total methane concentration at each sampling location is expressed as the sum of the dispersion contributions from all identified point sources and the background concentration. Based on this physical framework, the Emission-Partition model is formulated to jointly solve multiple unknown emission-related parameters. The mathematical expressions of the model are given in Equations 1–4 (Shi et al., 2023). σ i , y = a ⋅ x i b (1) σ i , z = c ⋅ x i d (2) F = ∑ k = 1 m C ′ m − C m C m 2 (3) C ′ m = ∑ i = 1 n q i ′ 2 π u ′ σ i , y ′ σ i , z ′ exp − y i ′ 2 σ i , y ′ 2 exp − z − H i ′ 2 σ i , z ′ 2 + exp − z + H i ′ 2 σ i , z ′ 2 + B ′ (4)

Where, Q i is the CH 4 emission intensity of the i point source, i = 1,2,3,…, n , n is the total number of strong point sources; u is the wind speed, H i is the effective emission height of the i point source, σ i,y and σ i,z are the horizontal and vertical dispersion parameters of the i point source, respectively, and B is the background CH 4 concentration in the measurement area. a and b are horizontal diffusion coefficients, and c and d are vertical diffusion coefficients. For the m th sampling location ( m = 1,2,3 , … , M , where M denotes the total number of concentration samples), C ( m ) represents the methane concentration directly measured by the UAV–AirCore system. During the inversion process, Equations 1, 2 together with Equation 4 constitute the forward physical model, which maps model parameters to simulated methane concentrations at the sampling locations. For each iteration of the optimization procedure, a candidate parameter set ( Q i , H i , a , b , c , d , B ) is generated. Based on this parameter set, the dispersion parameters are first calculated using Equations 1, 2, and the corresponding simulated concentration C ′ ( m ) is subsequently computed using Equation 4. Here, C ( m ) represents the methane concentration directly observed by the UAV–AirCore system and is treated as a fixed observational quantity, whereas C ′ ( m ) varies with the parameter combination. The spatial coordinates of the sampling location are denoted by ( x , y , z ) . Parameter inversion is achieved by minimizing the residuals between simulated and observed methane concentrations.

The Emission-Partition model adopts a physically constrained inversion framework, in which parameter estimation is achieved through two sequential optimization modules: particle swarm optimization (PSO) and the interior point penalty function (IPPF) method. The framework is built upon a multi-source Gaussian dispersion model and iteratively adjusts model parameters to achieve optimal agreement between simulated methane concentrations and UAV–AirCore observations.

In the initial stage, PSO is employed to perform a global search over the high-dimensional and nonlinear parameter space. Each candidate parameter set is treated as a particle, representing a complete combination of unknown parameters, including emission rate, effective release height, dispersion parameters, background methane concentration, and wind field variables. A swarm consisting of several tens of particles is adopted to ensure sufficient diversity of candidate solutions. Parameter updates are guided by both individual best solutions and the global optimum within the swarm. For each particle, methane concentrations at the UAV–AirCore sampling locations are simulated using the Gaussian dispersion Equations 1–4, and the discrepancy between simulated and observed concentrations is evaluated through an objective function. The PSO procedure is executed for 10,000 iterations, allowing adequate exploration of the solution space and reducing sensitivity to the initial parameter settings. During the PSO process, all parameters are restricted within predefined physically reasonable bounds to prevent non-physical solutions.

Following the global search, the IPPF method is applied to further refine the parameter estimates through constrained local optimization. In this stage, physical constraints on emission strength, dispersion parameters, background concentration, and wind field variables are explicitly incorporated by introducing penalty terms into the objective function. When a candidate solution approaches the predefined constraint boundaries, the penalty contribution increases progressively, resulting in a higher objective function value and discouraging boundary-adjacent solutions. If a parameter set exceeds the allowable range, the penalty term becomes dominant and forces the optimization trajectory back toward the interior of the feasible domain. As the iteration proceeds, the influence of the penalty terms is gradually strengthened, such that early iterations emphasize reducing the mismatch between simulated and observed methane concentrations, while later iterations increasingly enforce physical consistency. The IPPF-based local optimization typically converges within several hundred iterations, yielding stable and physically plausible parameter estimates.

As illustrated in Figure 2, the inversion procedure begins with defining reasonable bounds for all unknown parameters and constructing the objective function (Equation 3) to quantify the mismatch between simulated concentrations C ′ ( m ) and observations C ( m ) . The feasible parameter regions identified by PSO are subsequently used as constraints for the IPPF-based local optimization. This two-stage optimization strategy enables stable reconstruction of the methane concentration field and reliable estimation of emission rates and effective release heights for individual point sources. Methane concentration measurements acquired by the UAV–AirCore system, together with meteorological parameters and spatial location information, constitute the input dataset for the Emission-Partition model.

All calculations are implemented in Matlab (MathWorks) using custom-developed codes for both PSO-based global optimization and IPPF-based constrained local optimization. Model performance is evaluated by point-by-point comparison between simulated methane concentrations and UAV–AirCore observations using regression analysis and goodness-of-fit metrics. In addition, the emission rates retrieved by the Emission-Partition model are compared with estimates derived from several commonly used quantification approaches to assess the relative stability and applicability of different methods for methane emission estimation in oilfield environments.

To evaluate the accuracy of the Emission-Partition model in estimating methane emissions in oilfield areas, this study selected several mainstream point-source quantification methods for comparative analysis of methane emission rates from individual point sources. These methods include the OTM 33A model (Edie et al., 2020; Heltzel et al., 2020) proposed by the US EPA, the mass balance method (Cambaliza et al., 2014; Heimburger et al., 2017), NLSF (Shi et al., 2022), and the facility emissions calculation method NLSF (Mitchell et al., 2015).

It should be noted that the NLSF approach and the Emission-Partition model are based on the same physical framework, namely, a multi-source three-dimensional Gaussian dispersion model. The comparison between these two methods therefore focuses on differences in parameter inversion strategies rather than differences in the underlying physical assumptions. Specifically, NLSF employs a traditional gradient-based nonlinear optimization scheme, whereas the Emission-Partition model adopts a hybrid optimization framework combining PSO and the IPPF to improve solution stability under complex, multi-parameter conditions.

Multiple horizontal flight missions were conducted with the UAV–AirCore system to sample CH 4 concentrations throughout the study area and to identify anomalous methane distributions. In this study, anomalous regions are defined as spatially coherent areas characterized by CH 4 concentration enhancements that are significantly higher than the regional background, serving as spatial constraints for subsequent quantitative inversion and three-dimensional sampling.

To ensure the reliability of CH 4 emission quantification, two screening criteria were applied to the UAV–AirCore dataset (Tong et al., 2023; Zhu et al., 2024).

1. It is essential that wind speed, wind direction, humidity, and atmospheric pressure are successfully collected at a high frequency during UAV-AirCore data collection.

2. The mean wind speed during Aircore system data collection should be greater than 2.0 m/s.

The aforementioned two criteria were employed to select six sets of valid data from the entire sampling results, which were then used to demonstrate the areas of abnormal CH 4 concentration in the oilfield region.

The outcomes of the horizontal flight measurements conducted with the UAV-AirCore system and the Emission-Partition model simulations in the two study areas are presented in Figures 3, 4, respectively. The leftmost section of the image depicts the results of the Aircore system, which clearly demonstrate the anomalous area of CH 4 concentration. The image’s central section displays the outcomes of the Aircore measurements and the Emission-Partition model simulations. The blue dots represent the measured data obtained from the Aircore system, while the red dots illustrate the simulation results generated by the Emission-Partition model. The comparison demonstrates that the simulation results of the Emission-Partition model are in close agreement with the original measurement results, and that the trends of the two data sets are largely consistent. In order to facilitate a more detailed comparison of the two results, we proceeded to fit the two sets of data in question linearly. The results of this fitting are presented in the rightmost section of Figures 3, 4. The results indicate a high correlation between the two sets of data, with the R2 value of approximately 0.85. This demonstrates that the Emission-Partition model is capable of reconstructing CH 4 concentration with a reasonable degree of effectiveness.

After identifying methane concentration anomalies through horizontal flight surveys, top-down spiral flight measurements were conducted over selected hotspot areas using the UAV–AirCore system to further refine source localization and emission quantification. This flight strategy enables continuous vertical profiling of CH 4 concentrations from higher altitudes down to near-surface levels within a limited horizontal footprint, providing a three-dimensional characterization of concentration anomalies.

As shown in Figure 5, the spiral vertical flight measurements clearly reveal the vertical structure of methane concentrations. Distinct concentration enhancements were observed at altitudes of approximately 150 m in both study areas, indicating that the emission sources were associated with elevated effective release heights rather than near-surface emissions. Compared with horizontal or single-altitude measurements, vertical profiling substantially improves constraints on the altitude of concentration maxima, vertical gradients, and background concentration levels. For quantitative inversion, the vertically resolved CH 4 concentration data, together with simultaneously measured meteorological parameters (e.g., wind speed and wind direction), were jointly input into the Emission-Partition model. Physically reasonable upper and lower bounds were assigned to all unknown parameters, and a hybrid optimization scheme combining PSO and the IPPF was applied to jointly estimate emission intensity ( Q i ) , effective emission height ( H i ) , dispersion coefficients ( a – d ), background concentration ( B ) , and wind field parameters ( W s and W d ). The inclusion of vertical concentration profiles effectively reduced parameter non-uniqueness, particularly by constraining the coupling between Q i and H i , thereby enabling refined quantitative emission estimates.

Model accuracy was evaluated by comparing simulated and observed methane concentrations. As reported in Table 1, the reflection index reaches 0.97 for Area 1 and 0.98 for Area 2, indicating strong agreement between modeled and measured concentration fields and demonstrating the robustness of the inversion results. Uncertainty analysis shows that the uncertainties associated with the retrieved parameters are generally limited. The uncertainty in emission intensity is approximately ± 3.7 % for Area 1 ( 3.22 ± 0.12 g s − 1 ) and ± 1.1 % for Area 2 ( 3.68 ± 0.04 g s − 1 ), while the uncertainty in effective emission height ranges from ± 9.2 % for Area 1 ( 27.2 ± 2.5 m) to ± 9.0 % for Area 2 ( 35.4 ± 3.2 m). Uncertainties in corrected wind speed are within ± 0.3 m s − 1 for Area 1 and ± 0.2 m s − 1 for Area 2, while uncertainties in corrected wind direction are less than ± 1.5 ° and ± 0.8 °, respectively. The primary sources of uncertainty include (1) vertical heterogeneity of local wind fields, (2) atmospheric turbulence during the sampling period, and (3) temporal synchronization errors between AirCore sampling and positional data. Overall, the vertical flight strategy significantly enhances the stability and accuracy of methane emission inversion by providing stronger observational constraints on the vertical concentration structure.

The agreement between simulated and measured CH 4 concentrations presented in Sections 4.1, 4.2 demonstrates that the Emission-Partition model can reconstruct methane concentration fields in oilfield environments under given meteorological conditions. Based on these results, a comparative analysis was conducted to evaluate the performance and stability of the proposed approach in point-source methane emission quantification. Emission estimates obtained using the Emission-Partition framework were compared with results derived from four commonly used approaches, including the Mass Balance method, NLSF, the OTM 33A method, and facility-based emission inventory calculations.

These approaches are based on different data sources and calculation principles. The Emission-Partition model, NLSF, OTM 33A, and the Mass Balance method estimate methane emissions from observed concentration enhancements combined with meteorological information, using dispersion modeling or flux integration to infer emission rates. Facility-based inventory calculations derive emission estimates from engineering parameters and operational data of emission facilities, such as airflow rates and methane concentrations at outlets. In this study, inventory-based results provide an engineering reference for comparison of emission magnitudes.

Using UAV–AirCore observations, the Emission-Partition model retrieves emission rates, effective release heights, dispersion parameters, background concentration, and corrected wind field parameters by minimizing residuals between simulated and observed methane concentrations. The spatial distribution of inferred emission sources corresponds to the methane enhancement patterns identified during UAV measurements. Table 2 summarizes the locations and emission intensities of major emission sources identified in the monitoring area. The inversion results show that, within each study area, one emission source contributes most of the observed methane enhancement and controls the overall spatial structure and magnitude of the concentration field. This source is therefore identified as the dominant emission source based on inversion results and concentration field reconstruction.

Figure 6 presents a comparison of CH 4 emission intensities estimated using different methods for the two study areas. Emission estimates from all methods fall within the same order of magnitude, while differences appear in the distribution of results. The boxplots in Figure 6 describe the dispersion of emission estimates obtained from multiple valid datasets for each method. This dispersion reflects the consistency of emission estimates across datasets and provides information on method stability under varying observational and meteorological conditions. The OTM 33A method shows the largest dispersion of emission estimates, consistent with its sensitivity to wind field variability and transient plume structure. The Mass Balance method produces relatively lower emission estimates with moderate dispersion, influenced by plume interception and spatial averaging during flux integration. The NLSF method yields intermediate dispersion, corresponding to the behavior of gradient-based optimization in a multi-parameter Gaussian dispersion framework. The Emission-Partition model exhibits the smallest dispersion among the compared methods, indicating stable performance when three-dimensional UAV observations are combined with physically constrained hybrid optimization. Quantitative uncertainty was further evaluated using the standard error of emission intensity estimates under identical observational conditions, with results summarized in Table 3. The standard errors represent uncertainty associated with the inversion process, incorporating residuals between modeled and observed concentrations as well as parameter sensitivity.

As shown in Table 3, the Emission-Partition model yields the smallest standard errors in both study areas, followed by the NLSF method. Larger standard errors are obtained for the Mass Balance and OTM 33A methods, reflecting their sensitivity to wind variability, plume intermittency, and sampling configuration. Standard errors are not listed for the facility-based inventory because emission estimates are derived from deterministic engineering calculations based on facility operating parameters, and uncertainty propagation consistent with atmospheric inversion methods is not available.

Overall, the comparison indicates that the Emission-Partition model produces emission estimates consistent with engineering-based reference values and demonstrates improved stability and reduced uncertainty compared with other atmospheric observation–based approaches. The integration of three-dimensional UAV–AirCore measurements with physically constrained optimization provides a reliable framework for methane point-source quantification in complex oilfield environments.