UAV-Based Measurements of Methane Enhancements Reveal Hotspot Structure and Wind Effects

Abstract

Methane (CH₄) emissions from confined animal feeding operations are spatially heterogeneous, and uncrewed aerial vehicle (UAV)-based concentration measurements depend critically on atmospheric conditions. Yet the limits of spatial interpretability under varying wind regimes remain poorly characterized. This study integrates anisotropy-aware geostatistical analysis with atmospheric stability classification to evaluate the spatial interpretability of UAV-derived near-field methane enhancements (ΔCH₄) over a commercial feedlot. Grid-based UAV surveys under contrasting wind regimes show that weakly unstable conditions (wind speed ≳ 2 m s^–1, turbulence intensity ≲ 0.35) produce elongated, wind-aligned plume structures with strong directional coherence (D ² up to 69%; |Δφ| ≤ 13.5°) and low interpolation uncertainty (CV-RMSE = 0.062–0.101 ppm). Extremely unstable conditions yield fragmented, near-isotropic ΔCH₄ patterns with substantially higher uncertainty (CV-RMSE = 0.215–0.367 ppm). High directional coherence (D ² > 36%) coincided with persistent wind direction and low coherence (D ² < 12%) occurred exclusively under highly variable, extremely unstable conditions. These results demonstrate that the wind regime and atmospheric stability govern the spatial interpretability of UAV-derived ΔCH₄ fields, with direct implications for survey design and data quality assessment. Future work should incorporate onboard wind measurements, multialtitude sampling, and inverse dispersion modeling to enable quantitative flux estimation at the facility scale.

1. Introduction

Methane (CH₄) is a potent, short-lived greenhouse gas (GHG) with a global warming potential approximately 28–30 times greater than that of carbon dioxide over a 100-year time scale. Livestock production is one of the major anthropogenic CH₄ sources, responsible for 30–40% of global agricultural CH₄ emissions through ruminant enteric fermentation and manure management. Accurate characterization of CH₄ emissions from livestock systems is, therefore, critical for improving greenhouse gas inventories and mitigation strategies.

Characterizing CH₄ patterns at the facility scale remains technically challenging. Emission sources often operate simultaneously, vary in strength, and interact with rapidly evolving near-surface wind fields, complicating the interpretation of observed methane concentration enhancements above background (ΔCH₄). Conventional field-based techniques, such as chambers, open-path lasers, and flux towers, provide accurate local measurements but cannot resolve the lateral concentration gradients needed to distinguish emissions from adjacent features. , Instead, satellite- and aircraft-based CH₄ retrievals offer regional coverage but lack the spatial resolution (25 m or coarser) required to resolve fine-scale concentration structure within heterogeneous emission landscapes.

Recent advances in lightweight, high-precision CH₄ sensors have enabled uncrewed aerial vehicles (UAVs) to collect dense spatial CH₄ measurements along flexible flight paths, improving the ability to map fine-scale concentration variability compared with traditional ground- and tower-based approaches. , This high-density sampling bridges the observational gap between point-source ground sensors and the coarser resolutions typically provided by satellite or manned-aircraft platforms. While UAV platforms can resolve fine-scale ΔCH₄ variability, interpretation remains nontrivial because wind speed, wind direction, turbulence, atmospheric stability, surface roughness, and facility layout strongly influence plume geometry and persistence. Wind-driven advection and mixing can reshape, displace, or diffuse methane plumes prior to and during sampling, meaning that ΔCH₄ enhancements observed over specific surface features may reflect either local emissions or plumes advected from upwind sources. A key unresolved question is thus not only where ΔCH₄ enhancements occur but also under which atmospheric flow regimes they retain sufficient spatial coherence for reliable interpretation. Under some conditions, plume structure retains spatial coherence; under others, it degrades, limiting reliable source attribution. Without accounting for these wind-driven effects, near-field CH₄ concentration maps risk being overinterpreted regardless of measurement resolution.

Although UAV-based CH₄ mapping has advanced rapidly, most existing studies rely on visual inspection or simple interpolation to interpret plume patterns. These approaches rarely quantify plume orientation, anisotropy, or spatial coherence, properties that are strongly shaped by wind-driven dispersion. , Geostatistical tools, such as variogram analysis, anisotropy-aware kriging, and spatial hotspot statistics, provide the ability to formally characterize spatial dependence and plume structure, yet they have rarely been applied to near-field CH₄ measurements over heterogeneous emission environments. This methodological gap limits our ability to assess when and where ΔCH₄ patterns can be confidently attributed to surface sources.

Concentrated livestock operations, such as beef feedlots, provide an ideal test environment for developing and evaluating wind-informed spatial interpretation methods. Feedlots are characterized by open pen layouts, exposed manure surfaces, and high animal densities with methane generated primarily through enteric fermentation by the feedlot cattle and a smaller contribution from the anaerobic decomposition of manure. Enteric fermentation is the dominant source of CH₄ emissions from livestock systems, typically accounting for ∼80–90% of emissions, while manure management contributes a smaller but variable fraction. Emissions are spatially heterogeneous and are influenced by animal distribution, manure moisture content, pen management practices, and the presence of wastewater or runoff retention lagoons. Previous studies have demonstrated substantial temporal and spatial variability in CH₄ emissions from feedlots. , However, these studies have primarily emphasized emission rates or source strength estimation, with comparatively less emphasis on explicitly quantifying how atmospheric flow regimes shape plume morphology and affect spatial attribution of near-field concentration patterns within the facility. This complexity enables a robust assessment of how wind regimes influence plume coherence, spatial pattern, and attribution confidence across diverse emission structures.

To address these gaps, this study develops a wind-informed analytical framework that, to our knowledge, is among the first to explicitly integrate concurrent wind measurements and atmospheric stability classification into the geostatistical interpretation of UAV-based CH₄ measurements over a heterogeneous emission environment. Using repeated grid-based UAV surveys over a commercial feedlot, this study quantifies how wind speed, direction, and variability govern plume presence, spatial pattern, and directional coherence in ΔCH₄ fields. We apply geostatistical and directional analyses to characterize plume anisotropy, spatial dependence, and hotspot alignment, enabling a systematic evaluation of attribution confidence under contrasting wind regimes. By explicitly linking wind conditions to spatial observability in near-field methane measurements, this work establishes general constraints on the interpretation of high-resolution ΔCH₄ data with implications for source attribution and uncertainty characterization in facility-scale atmospheric methane studies.

2. Materials and Methods

Methane concentrations were measured during seven UAV flight missions conducted on 29 October 2024, 30 October 2024, and 22 August 2025 (Figure a). The flights were carried out at a commercial feedlot located in southern Alberta, Canada, situated within a predominantly flat agricultural prairie landscape characterized by low topographic relief and extensive surrounding cropland. The facility occupies approximately 65 ha (∼800 m × 800 m; Figure a), and houses approximately 27,000 beef cattle.

1.

(a) Grid Flight Plan over the study area; (b) UAV platform equipped with a MIRA Strato Series CH₄ sensor and forward air inlet for CH₄ measurements.

The feedlot consists of 109 production pens arranged in a rectangular grid configuration, 17 smaller pens designated for receiving and shipping animals, and five runoff lagoons located primarily along the southern and western boundaries of the site. Individual production pens have areas of approximately 0.03 ha (∼300 m²). Pens are separated by service alleys and feed bunks and are bordered primarily by fencing and low-height infrastructure. Vertical structures within the facility include fencing (∼1.2 m), feed bunks, light poles, two single-story office or service buildings, and localized feed storage structures near the northern boundary, including a vertical silo. These features occupy a small portion of the facility footprint, and the surrounding terrain is largely unobstructed agricultural land, minimizing terrain-induced flow distortion.

At this facility, manure accumulates within occupied pens, resulting in colocated enteric and manure emissions at the pen scale (Supporting Information, Figure S1). A DJI Matrice 300 RTK quadcopter UAV equipped with a MIRA Strato Series CH₄ gas analyzer (AERIS Technologies Inc., USA), a mid-infrared cavity-enhanced laser spectrometer, was used to collect atmospheric CH₄ data along preprogrammed grid flight paths (Supporting Information, SI Section 2.1, Table S1). The UAV was flown at a constant altitude of 20 m above ground level (AGL) to maintain safe clearance from overhead infrastructure, such as power lines, and to minimize interference with regular feedlot operations. A small subset of western pens was excluded from the grid flight coverage due to temporary operational constraints. Although a few of these pens had cattle in them, the wind direction during sampling did not usually blow from those pens toward the area where measurements were taken. Because of this, they likely had little effect on the ΔCH₄ patterns that were observed. Each sampling event consisted of a single autonomous grid flight lasting approximately 25 min and covering the accessible facility footprint. The UAV followed a full-area grid pattern directly over production pens and lagoons, enabling measurements both above active emission sources and across the downwind portions of the site. Horizontal source-receptor distances during transects ranged from near-zero (direct overflight of emitting pens) to several hundred meters, with individual transects spanning approximately 700–750 m across the facility. Given the UAV’s typical flight speed of 8–11 m s^–1 and 5 s measurement intervals, this corresponds to an effective spatial sampling distance of approximately 40–55 m along each transect, yielding on the order of 15–20 independent samples per pass. At the sampling height of 20 m AGL, observed ΔCH₄ enhancements primarily reflect near-field plume transport within the facility rather than long-range atmospheric dispersion.

The analyzer sampled ambient air through a forward-facing intake to reduce potential rotor-wash interference (Figure b). The sampling inlet consisted of a 12 mm diameter tube extending approximately 1 m forward from the UAV frame, placing the intake ∼250 mm ahead of the rotor plane. Rotor downwash can influence gas concentration measurements by disturbing plume structure or entraining ambient air, particularly when sampling near localized emission sources. However, experimental studies have shown that differences between body-mounted and extended-arm inlets become minimal when concentrations are averaged over several seconds or when sampling spatially mixed plumes.

Seven 25 min flights were conducted, two each on 29 October 2024 (12:00 and 2:00 p.m.), two on 30 October 2024 (9:30 and 11:00 a.m.), and three on 22 August 2025 (11:45 a.m., 12:30 p.m., and 2:00 p.m.). Flight records were uploaded to the AirData UAV platform, a cloud-based flight data management and analytics system, and 5 s wind vector estimates were retrieved using AirData’s proprietary wind algorithm, v2.2. This algorithm estimates horizontal wind speed and direction from onboard UAV telemetry, including GNSS-derived ground speed, aircraft attitude, and inertial measurements, by resolving the vector difference between the aircraft’s motion relative to the ground and the surrounding air mass.

Figure a illustrates the UAV flight plan over the study area and Figure b illustrates the UAV instrumentation and sampling configuration. The complete CH₄ mapping and analysis workflow, from data acquisition to geospatial synthesis, is summarized in Figure .

2.

Overview of the methodological framework used for UAV-based methane measurements, wind characterization, spatial interpolation, and hotspot analysis.

2.1. Methane Detection Sensor

The MIRA Strato Series analyzer is a mid-infrared laser spectrometer that employs cavity-enhanced absorption spectroscopy for real-time, high-precision gas detection. Each unit is factory-calibrated before deployment and provides 1 ppb precision for CH₄. Its fast response time and sensitivity enabled detection of fine-scale concentration gradients within feedlot plumes. Detailed technical specifications (measurement range, operating conditions, and noise performance) are provided in Supporting Information (SI Section 2.1, Table S1).

2.2. Data Processing

Methane concentrations were recorded at 1 s resolution, whereas wind measurements were available at 5-s intervals. To synchronize the data sets, CH₄ observations were aggregated to 5-s intervals prior to merging with the wind data. Wind data were not interpolated to higher temporal resolution to avoid introducing artificial variability and to ensure pairing with the observed wind vectors. Measurements collected during UAV turning maneuvers at the ends of grid transects were excluded to avoid positional clustering and airflow artifacts near the sampling inlet. Only observations from straight flight segments were retained, and CH₄ records without a corresponding wind measurement during temporal matching were excluded.

To calculate ΔCH₄ by minimizing the regional background levels in the feedlot environment, we applied a flight-specific background correction to all CH₄ concentration data. For each UAV flight, the 10th percentile of measured CH₄ concentrations was defined as the background level. The ΔCH₄ was calculated using eq

$${\mathrm{\Delta }\mathrm{CH}}_4={\mathrm{CH}}_{4,\mathrm{measured}}-{\mathrm{CH}}_{4,\mathrm{background}(\mathrm{10th\; percentile})}$$ 1

To evaluate the stability of background CH₄ concentrations and the robustness of the percentile-based background correction, additional background measurements were collected during dedicated UAV flights over cropland adjacent to the feedlot, representing a low-emission environment. Analysis of these clean-air segments showed CH₄ variability, typically between 0.01 and 0.03 ppm, with a maximum of 0.08 ppm in one segment, well below the plume-attributed enhancements of 0.1–1 ppm observed during flights in a grid pattern. Background CH₄ concentrations for the seven flights in a grid pattern were also recalculated using the fifth, 10th, and 15th percentiles of the CH₄ distributions within each flight. The resulting background values differed by less than 0.03 ppm across flights, indicating that the estimated background was not sensitive to the exact percentile threshold. The 10th percentile was therefore retained as a reasonable flight-specific background estimate, consistent with approaches used in previous atmospheric methane analyses.

For spatial analysis, a uniform raster grid covering the flight path was generated, in which the grid resolution was defined as half of the median nearest-neighbor distance between sampled points, calculated using the Fast Nearest Neighbor (FNN) package (version 1.1.3) in R (version 4.3.1) to avoid oversmoothing. The resulting grid had a final resolution of 17 × 17 m and was projected in UTM Zone 12N with metric units.

Given that the monitored feedlot covered approximately 65 ha (∼800 × 800 m; Figure a), the 17 m grid spacing corresponds to ∼2% of the facility width and yields ∼45–50 cells along each horizontal axis. This resolution is well below the spatial correlation ranges observed in the variograms (45–150 m; Table S2), supporting the resolution of facility-scale ΔCH₄ gradients without excessive smoothing.

2.3. Wind Characterization

The wind data from each UAV flight was used to calculate the circular mean wind direction (θ̅), mean wind speed (U̅), standard deviation of wind speed (σ_u), coefficient of variation (CV = σ_u/U̅), equivalent to the turbulence intensity (TI) commonly used in boundary-layer meteorology, and circular standard deviation (σ_θ ), following the formulations of Zar (1999) and Berens (2009). The circular mean direction (θ̅), resultant vector length (R), and circular standard deviation (σ_θ ) were calculated as shown in eqs –, respectively

$$\mathrm{\theta ̅}={\tan }^{-1}(\frac{1}{n}\sum _i\sin {\theta }_i,\frac{1}{n}\sum _i\cos {\theta }_i)$$ 2

$$R=\sqrt{{(\overline{\cos \theta })}^2+{(\overline{\sin \theta })}^2}$$ 3

$${\sigma }_{\theta }=\sqrt{-2\ln R}\times \frac{180}{\pi }$$ 4

All angles were expressed in meteorological convention (0° = wind from the north, increasing clockwise). These wind statistics were used for interpreting CH₄ plume orientation and to parametrize anisotropy and directional hotspot analyses.

To relate meteorological variability to plume structure, each flight was also categorized into qualitative flow regimes using U̅, σ_θ , and TI. Following the conceptual framework of regime-based analyses of boundary-layer turbulence and the observed relationships between wind direction variability and mean wind speed, three flow regimes were defined: (i) steady/coherent flow (U̅ > 5 m s^–1, σ_θ < 15°, TI < 0.2); (ii) moderately variable flow (2 ≤ U̅ ≤ 5 m s^–1, 15° ≤ σ_θ ≤ 40°, 0.2 ≤ TI ≤ 0.4); and (iii) calm/variable flow (U̅ < 2 m s^–1, σ_θ > 40°, TI > 0.4). When metrics disagreed, regimes were assigned by majority across (U̅, σ_θ , TI), and ties were resolved toward the more variable class (Calm > Moderately variable > Steady).

2.4. Validation of UAV-Derived Wind Measurements

The reliability of the UAV-derived wind measurements was evaluated using two independent approaches. In the first approach, a three-dimensional (3D) sonic anemometer installed at 7 m above ground level (AGL) on a meteorological tower within the feedlot during the UAV-CH₄ measurement campaign served as the reference wind measurement system. Sonic wind speed and direction were available as 15 min averages. For each UAV grid flight, flight start and end times were used to identify overlapping sonic measurement intervals, and the temporal overlap between each 15 min sonic interval and the UAV flight window was used as a weighting factor to derive flight-matched sonic wind conditions. The 15 min averaging interval reflects the temporal resolution of the sonic anemometer output and provides stable estimates of mean wind conditions. Wind observations were decomposed into orthogonal horizontal components, and sonic wind speed was calculated as the magnitude of the time-weighted vector mean, while wind direction was derived from the corresponding vector-averaged components to correctly account for angular wrap-around. Surface roughness length (z ₀) and Monin–Obukhov length (L) were derived from turbulence statistics and flux calculations based on high-frequency measurements from the 3D sonic anemometer. Missing z ₀ values, indicating invalid values from the turbulence-based roughness calculation, occurred in two 15 min intervals and were replaced by linear interpolation from adjacent valid observations within the same day prior to averaging. Atmospheric stability for each flight was characterized using the dimensionless stability parameter ζ = z/L, evaluated at the UAV operating height (z = 20 m), and assigned to a stability class following the scheme of Sorbjan and Grachev (2010).

Because UAV wind measurements were obtained at approximately 20 m above ground level (AGL), sonic wind speeds measured at 7 m were vertically extrapolated to 20 m by using three commonly applied surface-layer wind profile formulations. First, the empirical power law was applied as shown in eq .

$$U(z)=U(z_1){\left(\frac{z}{z_1}\right)}^{\alpha }$$ 5

where U(z) is the wind speed at height z, U(z ₁) is the measured wind speed at reference height z ₁, and α is the wind shear exponent. A value of α = 0.14 was adopted, which is commonly used for near-neutral atmospheric conditions over moderately rough terrain.

Second, the neutral logarithmic wind profile was used as shown in eq ,

$$U(z)=U(z_1)\frac{\ln (z/z_0)}{\ln (z_1/z_0)}$$ 6

where z ₀ is the rough surface length.

Finally, a stability-corrected logarithmic profile based on Monin–Obukhov similarity theory (MOST) was applied to account for atmospheric stability effects as shown in eq ,

$$U(z_2)=U(z_1)\frac{\ln (z_2/z_0)-\psi (z_2/L)+\psi (z_0/L)}{\ln (z_1/z_0)-\psi (z_1/L)+\psi (z_0/L)}$$ 7

where L is the Monin–Obukhov length and ψ represents the stability correction function. Stability corrections were applied using the Businger–Dyer flux-profile relationships for stable and unstable atmospheric conditions. Wind direction was assumed to remain constant with height within the surface layer, consistent with classical surface-layer similarity theory. UAV wind measurements from individual flights in a grid pattern were then summarized for each flight, with the mean wind speed calculated as the arithmetic mean of scalar wind speed observations and wind direction calculated from vector-averaged horizontal wind components. We acknowledge that the feedlot’s heterogeneous surface and limited fetch may support internal boundary layer development above approximately 5 m AGL, potentially causing the true wind profile to deviate from standard similarity relationships and introducing additional uncertainty into the extrapolated reference wind speeds.

In the second approach, an independent field validation study was conducted at an open agricultural site near Lethbridge, Alberta, Canada, using the same UAV platform and AirData wind retrieval algorithm (v2.2) together with a calibrated two-dimensional (2D) sonic anemometer at 7 m AGL. The UAV was flown at the same height in repeated square transects around the tower to obtain colocated wind observations during three campaign days in October and November 2025. For both approaches, UAV-derived wind speed and direction were compared against the respective reference measurements by calculating mean bias error (MBE), root-mean-square error (RMSE), mean absolute error (MAE), relative bias for wind speed, and mean absolute error for wind direction.

2.5. Variogram Modeling and Anisotropy Detection

The spatial structure of ΔCH₄ for each UAV flight was characterized by using ordinary kriging (OK). Empirical variograms were constructed and fitted with an isotropic exponential model to estimate the nugget, sill, and range (c ₀, c, and α) under physically constrained parameter bounds (details of variogram computation, model fitting, and cross-validation procedures are given in Supporting Information Section S2.4.1–S2.4.3). Model performance was evaluated using leave-one-out cross-validation (LOOCV), and the resulting cross-validated root-mean-square error (CV-RMSE) was used to select the best-fitting structure. Directional dependence was assessed by testing geometric anisotropy informed by the mean wind direction. A series of candidate anisotropic models with varying orientations and range ratios were compared against the isotropic baseline. During anisotropy testing, variogram parameters from the isotropic fit (c ₀, c, α) were held constant, and only geometric anisotropy parameters (ϕ, b/a) were adjusted, following the approach of Goovaerts (1997) and Webster and Oliver (2007). An anisotropic model was retained only when it improved cross-validated accuracy and met geometric constraints. The final variogram, anisotropic when supported and isotropic otherwise, was used to generate kriged ΔCH₄ surfaces and associated prediction variances. Full model parameters, computational settings, empirical variogram construction, model fitting, cross-validation procedures, and anisotropy testing details are provided in the Supporting Information (Sections S2.4.1–S2.4.5).

2.6. Hotspot Identification and Directional Dependence Analysis

This analysis was designed to quantify whether spatially coherent ΔCH₄ enhancements exhibit systematic directional dependence with respect to ambient wind forcing. Hotspots were first identified as localized regions of elevated methane concentrations and subsequently evaluated for angular coherence relative to the prevailing wind directions.

To assess the wind influence, two complementary approaches were applied. First, a descriptive wind-weighted sectoral analysis was used to summarize the angular distribution of hotspots and their correspondence with the wind frequency. Second, formal directional statistical models were employed to test whether observed hotspot-wind alignment reflects systematic advective transport rather than random spatial variability. These methods are described in Sections 2.6.1 and 2.6.2, respectively.

2.6.1. Hotspot Delineation and Wind-Weighted Sectoral Analysis

Spatial clusters of ΔCH₄ were identified using the Getis–Ord Gi* statistic applied to kriged ΔCH₄ raster for each UAV flight. The Gi* statistic produces standardized z-scores indicating local spatial clustering. An 8-neighbor (queen) contiguity kernel was used. Statistical significance was initially evaluated using the Benjamini–Hochberg false-discovery rate (FDR = 0.05). Because strict FDR control produced very few significant pixels in several flights, hotspot patterns were also examined using a relaxed z-score threshold (z ≥ 1.645; p < 0.10) for exploratory spatial interpretation, consistent with commonly used Gi* hotspot classification thresholds. Hotspot and wind directions were each partitioned into 22.5° azimuthal sectors (θ) following standard practice for pollution roses using the 16 cardinal sectors (22.5° arcs). For each flight, the sectoral hotspot share H_θ (fraction of pixels within each sector that are classified as hotspots) and the wind-frequency distribution f_θ were calculated from the same angular bins. A wind-weighted hotspot alignment index (A) was introduced as shown in eq ,

$$A=\sum _{\theta =1}^{16}{H}_{\theta }{f}_{\theta }$$ 8

where θ indexes the 16 azimuthal sectors (22.5° bins). Because both H_θ and f_θ are normalized sectoral proportions (i.e., each sum to unity across θ), A ranges from 0 (no directional correspondence, diffuse or turbulent dispersion) to 1 (perfect overlap, strong advective alignment). This formulation represents the normalized sectoral co-occurrence of hotspots and winds, with no cosine or angular weighting applied. The complementary (C) value, C = 1 – A represents directional mismatch between hotspot and wind sectors. Conceptually, A is a new alignment metric, analogous to sector-weighted wind exposure indices used in wave-forcing studies, but it quantifies directional co-occurrence rather than mechanical forcing or energy flux. Building on this framework, we also considered wind-conditioned methods, such as direction-weighted means, the Conditional Probability Function (CPF), Conditional Bivariate Probability Function (CBPF), and bivariate polar plots, which motivate our focus on directional relationships but do not yield a single scalar measure of alignment like A. To complement the scalar metric, A, pollution-rose-style polar plots of H_θ and f_θ were used to visualize the alignment of hotspots relative to the mean downwind axis as shown in Figure .

4.

Directional distribution of methane hotspot share (H_θ ) and wind frequency (f_θ ) during all UAV flights. Blue bars show the sectoral hotspot share (H_θ ), defined as the fraction of pixels within each directional sector that are classified as hotspots, while orange bars show the wind-frequency distribution (f_θ ). The dashed line indicates the downwind direction relative to the measured wind direction for each flight.

2.6.2. Directional Statistical Modeling

To quantify the statistical strength and orientation of hotspot-wind coupling, two regression models, a Generalized Linear Model (GLM) and an Ordinary Least Squares model (OLS) were fitted for each flight. The GLM modeled hotspot occurrence (H_i ) as a function of pixel bearing, while the OLS model evaluated directional variation in kriged methane enhancement $(\mathrm{\Delta }\mathrm{C}{\mathrm{H}}_{4,i})$ . Both models used angular predictors cos θ_i and sin θ_i , where θ_i denotes each pixel’s bearing from the feedlot centroid. , Representing the direction with sine and cosine terms allows circular variables to be incorporated into linear models without discontinuity at 0°/360°. Further, a nonparametric Spearman correlation test was used to verify the robustness of these directional relationships as described below.

2.6.2.1. Generalized Linear Model (GLM)

Hotspot occurrence was modeled using a binomial GLM with a logit link as shown in eq ,

$$\mathrm{logit}[P(H_i=1)]={\beta }_0+{\beta }_{\cos }\cos {\theta }_i+{\beta }_{\sin }\sin {\theta }_i+{\beta }_W{\mathrm{W̅}}_i$$ 9

where P (H_i = 1) denotes the probability that pixel i is classified as a hotspot (i.e., H _i = 1), and 0 otherwise, β ₀ is the intercept, β _cos and β _sin are directional coefficients for the cosine and sine of bearing (θ_i ), and β_W represents the effect of mean wind speed (W̅ _i ). The logit link transforms the probability of hotspot occurrence into log-odds, allowing binary outcomes (hotspot = 1, nonhotspot = 0) to be modeled using a linear combination of predictors. Then, by using the directional coefficients (β _cos,β _sin), the preferred hotspot direction representing the orientation of maximum hotspot probability predicted by the model $(\mathrm{\varphi ̂})$ and its angular offset from the downwind axis were calculated using eqs – respectively

$$\mathrm{\varphi ̂}=a\tan 2({\beta }_{\sin },{\beta }_{\cos })$$ 10

$$\mathrm{\Delta }\varphi =\mathrm{\varphi ̂}-({\theta }_{\mathrm{wind}}+{180}^{◦})$$ 11

After model fitting, the deviance explained (D ²) was used as a standard measure of the goodness-of-fit for the binomial GLMs. This metric quantified the proportion of spatial variation in hotspot probability captured by the directional terms, where higher D ² and smaller Δϕ values indicate more coherent, advective plume transport. Parametric bootstraps (2,000 draws) were used to obtain 95% confidence intervals for $\mathrm{\varphi ̂}$ .

2.6.2.2. Ordinary Least-Squares Model (OLS)

To examine directional variation in CH₄ enhancement magnitude, an OLS model was applied using eq

$$\mathrm{\Delta }\mathrm{C}{\mathrm{H}}_{4,i}={\alpha }_0+{\alpha }_{\cos }\cos {\theta }_i+{\alpha }_{\sin }\sin {\theta }_i+{\epsilon }_i$$ 12

where ΔCH_4,i is the kriged methane enhancement at pixel i; α ₀ is the intercept; α _cos and α _sin are directional coefficients for the cosine and sine of bearing (θ_i ); and ε_i is the residual error term. The preferred direction of maximum enhancement $(\mathrm{\varphi ̂})$ was derived from the model’s directional coefficients (α _cos, α _sin) following the same formulation described in eqs –. , The model’s goodness-of-fit was evaluated by R ².

2.6.2.3. Nonparametric Correlation Test

For each flight, the sector-level Spearman correlation (ρ) between hotspot share H_θ and wind-sector frequency f_θ was computed. Absolute |ρ| values were used as a nonparametric measure of directional agreement.

2.7. Feedlot-Scale Visualization and Zonal Statistics

Feedlot boundaries were delineated from high-resolution satellite imagery using a geospatial Segment Anything Model v2 (SAM2) workflow, optimized for remote-sensing applications. The resulting polygons were manually verified and classified as lagoons or pens, and pens were further categorized as occupied or empty based on simultaneous animal inventory records. These feature polygons define the zones used for the subsequent spatial analysis. The analytical outputs from the preceding sections, including the kriged ΔCH₄ raster and directional hotspot metrics, were spatially integrated with the feedlot polygons to assess how ΔCH₄ varied across these structural features. Zonal statistics (mean, 95th percentile, and maximum ΔCH₄) were extracted from the kriged raster for each polygon class to quantify methane enhancement and hotspot intensity. These polygon-level metrics provided the basis for comparing relative ΔCH₄ levels among occupied pens, empty pens, and lagoons, thereby linking emission patterns to the facility layout.

3. Results and Discussion

3.1. Performance of UAV-Derived Wind Measurements

Across the seven feedlot flights in a grid pattern, the UAV-derived wind speed at 20 m showed a systematic positive bias relative to the MOST-corrected sonic reference, with overall performance metrics of MBE = +0.78 m s^–1, RMSE = 1.01 m s^–1, and MAE = 0.81 m s^–1, corresponding to an overall relative bias of +28% (Table ).

1. Flight-by-Flight Comparison of UAV and Sonic Wind Measurements during Feedlot Flights in a Grid Pattern,,

Date and Time	Sonic WS (m s –1 ) (20 m)	UAV WS (m s –1 ) (20 m)	Diff WS	Rel. Bias (%)	Sonic WD	UAV WD	Diff WD	Stability ζ
29 October 2024 12:00 p.m.	0.52	0.88	0.37	71	178	243	65	–2.78 (xu)
29 October 2024 02:00 p.m.	1.75	2.37	0.61	35	119	132	13	–2.74 (xu)
30 October 2024 9:30 a.m.	4.43	5.79	1.37	31	191	211	20	–0.10 (wu)
30 October 2024 11:00 a.m.	8.36	10.35	1.99	24	215	230	15	–0.04 (wu)
22 August 2025 11:45 a.m.	1.05	1.74	0.7	67	37	53	17	–3.92 (xu)
22 August 2025 12:30 p.m.	1.87	1.78	–0.09	–5	85	84	–2	–1.92 (xu)
22 August 2025 02:00 p.m.	1.55	2.07	0.52	34	68	124	56	–2.67 (xu)

^aDiff WS and Diff WD denote UAV-sonic differences in wind speed and direction, respectively.

^b Rel. Bias represents the relative bias of UAV wind speed relative to sonic measurements.

^cAtmospheric stability is expressed as ζ = z/L (xu = extremely unstable; wu = weakly unstable).

Relative bias decreased with increasing wind speed, declining from +71% at the lowest observed sonic wind speed (0.52 m s^–1) to +24% at the highest (8.36 m s^–1), indicating improved agreement under stronger and more steady wind conditions with reduced directional variability. The only exception occurred during the 22 August 12:30 flight (bias = −0.09 m s^–1, −5%), which coincided with the highest surface roughness length among all flights (z ₀ = 0.60 m) and the least unstable atmospheric conditions among the August flights (ζ = −1.92), resulting in a smaller vertical extrapolation correction and near-zero net bias.

Among the three vertical extrapolation methods, UAV-sonic differences were the smallest when using the neutral log-law profile (RMSE = 0.92 m s^–1), intermediate for the MOST-corrected profile (RMSE = 1.01 m s^–1), and largest for the power-law profile (RMSE = 1.29 m s^–1), reflecting the sensitivity of the feedlot comparison to extrapolation assumptions rather than a definitive ranking of profile accuracy. The comparatively weaker performance of the MOST correction likely reflects limitations of the Businger–Dyer stability functions under strongly unstable conditions (ζ < −1), which characterized five of the seven flights. The systematic positive wind speed offset is attributed to two compounding factors: spatial sampling differences between the UAV grid-flight area average and the sonic point measurement, and height-extrapolation uncertainty arising from the 4-fold variation in surface roughness length across flights (z ₀ = 0.15–0.60 m), reflecting the heterogeneous feedlot surface.

Wind direction agreement depended strongly on the prevailing wind speed. For the four flights with stronger and more directionally persistent flow (29 Oct 2 p.m., 30 Oct 9:30 a.m., 30 Oct 11:00 a.m., and 22 Aug 12:30 p.m.; mean sonic wind speeds 1.75–8.36 m s^–1), UAV-derived wind direction agreed with the sonic reference within ±20° (mean absolute error ≈ 13°; Table ). The three remaining flights, 29 Oct 12:00 p.m. (0.52 m s^–1), 22 Aug 11:45 a.m. (1.05 m s^–1), and 22 Aug 2:00 p.m. (1.55 m s^–1), showed more variable direction errors (17–65°), consistent with the inherent directional variability of weak or gusty flow rather than algorithm error. UAV-derived wind direction was also systematically offset clockwise relative to the sonic reference in six of the seven flights, suggesting a possible yaw bias caused by aerodynamic flow distortion from the payload mounted to the UAV during concentration measurements. Despite this offset, the UAV-derived direction correctly identifies the wind sector under moderate to strong flow and is sufficient for wind regime classification.

The height-matched Lethbridge validation provides a more direct test of algorithm accuracy, as it eliminates vertical extrapolation and spatial sampling as confounding factors. UAV-derived wind speed showed near-zero mean bias (MBE < 0.3 m s^–1, RMSE ≈ 1.1 m s^–1, MAE ≈ 0.85 m s^–1) across approximately 786 paired 5-s observations, with about 70% of estimates within ±1 m s^–1 of the reference.

Because this comparison eliminates both vertical extrapolation and spatial sampling differences, it confirms that the AirData algorithm itself produces reliable wind speed estimates. The positive bias observed in the feedlot comparison therefore likely arises from height extrapolation and spatial sampling differences rather than the intrinsic limitations of the wind retrieval algorithm. Wind direction errors in the field validation ranged from 17.8° to 22.2° depending on flight orientation, consistent with the feedlot results under moderate to strong wind conditions. On this basis, UAV-derived wind speed, direction, turbulence intensity, and circular standard deviation are considered sufficiently reliable for classifying wind regime and interpreting plume patterns at the spatial resolution of this study as presented in Sections and .

3.2. Wind Regimes, Atmospheric Stability, and Spatial Anisotropy of ΔCH₄

The two stability regimes identified in Section , weakly unstable (wu, ζ = −0.04 and −0.10) for the 30 October flights and extremely unstable (xu, ζ < −1) for all five remaining flights, corresponded to markedly different spatial structures in the ΔCH₄ field (Figure , Figure S2). Across all seven UAV flights, ΔCH₄ fields exhibited a clear spatial structure, but the geometry and coherence of methane enrichment depended strongly on prevailing wind conditions (Table ) and corresponding variogram anisotropy metrics (Table ). Flights characterized by coherent or moderately persistent winds (Table ) tended to show anisotropic elongation of the ΔCH₄ field (b/a ≈ 0.30–0.45), whereas calm or highly variable winds produced inconsistent anisotropy (Table ). However, b/a values overlapped across wind categories, reflecting the influence of multiple factors on variogram shape beyond the wind regime alone, including source geometry and domain-scale flow heterogeneity. Accordingly, b/a is interpreted here as a spatial prediction tool rather than a definitive indicator of the wind regime. Near-neutral stratification favors two-component axisymmetric turbulence with coherent shear-driven structure, whereas extreme instability promotes more isotropic turbulence states with suppressed directional coherence, providing a mechanistic basis for the contrasting ΔCH₄ field anisotropy observed between the wu and xu flights.

5.

Spatial mosaics of anisotropy-corrected ΔCH₄ (ppm) fields for UAV flights on 29–30 October 2024 (top panels) and 22 August 2025 (bottom panels), overlaid with segmentation masks delineating occupied pens (black), empty pens (gray), and lagoons (orange). Red arrows indicate wind direction corresponding to the wind speed and direction values shown above each panel. Maps show only the surveyed portion of the facility corresponding to UAV flight coverage, with some western pens excluded due to operational constraints. The 100 m scale bar applies to all panels.

2. Mean Wind Statistics, Variability, Directional Spread, and Stability for All UAV Flight Campaigns.

Date and time	Number of data points ( n )	Mean wind speed (m s –1 )	SD wind speed (m s –1 )	TI	Mean wind direction (°)	Circular SD (°)	Flow regime	L	ζ = z / L (Stability class)
29 October 2024 12 p.m.	244	0.83	0.35	0.42	237.0	46.3	Calm/variable	–7.2	–2.78 (xu)
29 October 2024 2 p.m.	284	2.34	0.75	0.32	133.6	28.1	Moderately variable	–7.3	–2.74 (xu)
30 October 20249:30 a.m.	276	5.79	0.87	0.15	210.7	8.3	Steady/coherent	–205	–0.10 (wu)
30 October 2024 11 a.m.	277	10.29	1.40	0.14	230.3	5.4	Steady/coherent	–531	–0.04 (wu)
22 August 2025 11:45 a.m.	184	1.74	0.80	0.46	47.0	73.8	Calm/variable	–5.1	–3.92 (xu)
22 August 2025 12:30 p.m.	190	1.78	0.84	0.47	86.0	39.1	Calm/variable	–10.4	–1.92 (xu)
22 August 2025 2 p.m.	169	2.07	0.87	0.42	122.1	63.1	Calm/variable	–7.5	–2.67 (xu)

3. Variogram Model Type, Cross-Validation (CV) Results, and Anisotropy Parameters for Each UAV Flight.

Date and time	Use anisotropy	Major direction, θa (°)	b / a	CV-RMSE (isotropic)	CV-RMSE (best)	ΔRMSE (%)
29 October 2024 12 p.m.	Yes	78.0	0.35	0.384	0.367	4.5
29 October 2024 2 p.m.	Yes	86.4	0.30	0.240	0.215	10.4
30 October 2024 9:30 a.m.	Yes	79.3	0.45	0.106	0.101	4.9
30 October 2024 11 a.m.	No (isotropic)	-	-	0.0623	0.0623	0.0
22 August 202511:45 a.m.	Yes	218.0	0.98	0.364	0.217	40.3
22 August 2025 12:30 p.m.	Yes	264.0	0.30	0.346	0.326	5.7
22 August 2025 2 p.m.	Yes	87.9	0.30	0.367	0.335	8.8

This wind dependence is also evident in the empirical variogram anisotropy ellipses (Figure ), which shift from near-isotropic geometry under calm/variable winds to a strongly elongated, wind-aligned structure under steady/coherent flow. It should be noted that some calm or variable wind flights also produced relatively low b/a ratios; however, in these cases the anisotropy reflects spatially concentrated near-source accumulation under light winds rather than coherent downwind elongation. These distinct processes are differentiated using the directional metrics |Δφ|, D ², and A. With the exception of the 30 October 2024 (11:00 a.m.) flight, where exceptionally high wind speeds enhanced turbulent mixing and dilution of the concentration field despite coherent transport (discussed below), ΔCH₄ fields during three of the four flights conducted on 29–30 October 2024 (29 October 12:00 p.m., 29 October 2:00 p.m., and 30 October 9:30 a.m.) showed pronounced elongation (b/a = 0.30–0.45). Specifically, during the 29 October (2:00 p.m.; xu, ζ = −2.74) and 30 October 2024 (9:30 a.m.; wu, ζ = −0.10) flights, major axes were offset from the mean wind direction by approximately 47° to 49° (Tables and ), indicating that while wind plays a primary role in shaping the ΔCH₄ field, moderate directional persistence can support plume elongation even under extremely unstable (xu) conditions, though the axis of maximum spatial continuity does not perfectly align with the mean wind vector. Because variogram anisotropy defines an axis rather than a vector, alignment with the downwind direction was evaluated using the minimum angular separation modulo 180°. In contrast, the 29 October 2024 (12:00 p.m.) flight, classified as calm/variable under extremely unstable conditions (xu, ζ = −2.78), exhibited reduced directional coherence and weaker spatial organization with a variogram reflecting more localized, less oriented correlation. Incorporating geometric anisotropy yielded modest but repeatable improvements in cross-validated prediction error (4–10% reduction in CV-RMSE), demonstrating that wind-aligned spatial continuity could be a true feature of the ΔCH₄ fields. The two weakly unstable (wu) flights achieved the lowest absolute interpolation uncertainty, with CV-RMSE of 0.062 ppm (30 October 11:00 a.m.) and 0.101 ppm (30 October 9:30 a.m.), compared to 0.215–0.367 ppm across the five extremely unstable (xu) flights, reflecting the greater spatial predictability of ΔCH₄ fields under near-neutral, shear-driven conditions. These findings are consistent with previous UAV- and aircraft-based CH₄ studies showing that coherent winds impose directional structure on concentration fields. ,

3.

Wind regimes and spatial anisotropy of ΔCH₄ fields during seven UAV flight periods. Left panels show wind roses derived from 5-s meteorological records grouped by wind-speed classes (m s^–1). Mean wind direction and circular standard deviation are reported below each plot (Mean ± circ SD). Right panels show anisotropy ellipses fitted to empirical variograms of ΔCH₄ enhancement. The red arrow indicates the major axis of spatial continuity (θ_a), measured clockwise from north, while the gray dashed arrow indicates the mean downwind direction. Ellipse elongation (b/a) quantifies anisotropy, with lower b/a values indicating stronger directional structure and b/a = 1 representing isotropic spatial correlation.

Contrarily, the August 2025 flights (Figure ), all conducted under extremely unstable conditions (xu; ζ ranging from −1.92 to −3.92) combined with calm or highly variable winds, produced less coherent plume structures. Wind roses show large directional spread (circ SD 39–74°) and low mean wind speeds (<2.1 m s^–1), consistent with conditions that favor intermittent plume transport rather than a stable advective plume. Anisotropic models still improved CV-RMSE (6–40%) but fitted anisotropy ratios ranged from nearly isotropic (b/a = 0.98) to moderately elongated (b/a = 0.30), and major-axis orientations varied widely (Table ). The 11:45 a.m. flight, showing the largest CV-RMSE reduction despite a nearly circular ellipse, illustrates how anisotropic models can capture subtle directional tendencies or local nonstationarity even when global plume elongation is weak. Similarly, the 2:00 p.m. flight produced a moderately elongated variogram (b/a = 0.30, ΔRMSE = 8.8%) despite a mean wind speed of only 2.07 m s^–1 and high directional spread (Circular SD = 63.1°). The major axis orientation (87.9°) deviated approximately 34° from the mean wind direction, suggesting that the observed spatial structure reflects a brief coherent transport episode within the flight window rather than the time-averaged flow, a pattern consistent with intermittent plume advection under variable winds rather than a missed or shifted plume. This pattern is consistent with expectations from advection-dispersion behavior and recent UAV evidence linking plume geometry to wind coherence and gust statistics. ,

A significant exception occurred on 30 October 2024 at 11:00 a.m. (Figure ). Despite being classified as steady/coherent with the highest recorded wind speeds (>10 m s^–1), it generated a ΔCH₄ field best represented by an isotropic variogram (b/a = 1.00). In this case, ΔCH₄ enhancements remained detectable across the domain, but spatial correlation became nearly uniform and incorporating anisotropy provided no predictive benefit. It should be noted that the variogram anisotropy ratio (b/a) and the hotspot directional metrics (D ², |Δφ|) measure fundamentally different aspects of the spatial pattern, b/a captures the geometric elongation of the continuous concentration field, while D ² and |Δφ| measure the directional clustering of statistically significant hotspot locations. Under the near-neutral (wu, ζ = −0.04) conditions of this flight, coherent shear-driven turbulence maintained directional transport and hotspot alignment consistent with Gucci et al. (2025), while the exceptionally high wind speed likely enhanced turbulent mixing and dilution of the concentration field, producing isotropic variogram geometry despite coherent hotspot directionality. This result highlights that sufficiently strong advection could homogenize the concentration structure, effectively erasing directional spatial signatures even when emissions are present. It is equally possible that at such high wind speeds (>10 m s^–1), the concentrated plume core remained vertically compressed below the UAV’s flight altitude or was transported beyond the sampling domain before it could laterally expand. In either case, while ΔCH₄ remained detectable (Table ), the lack of clear spatial structure suggests that the “sampling window” may need to be extended further downwind in future high-wind scenarios to capture the developed plume geometry. An anisotropic variogram was used only if leave-one-out cross-validation improved on the isotropic baseline rather than defaulting to the mean downwind direction. This data-driven approach aligns with recent CH₄ monitoring studies emphasizing empirical justification and flexible spatial models for advective fields. , Because anisotropy can arise from source layout or flight geometry, we interpret variogram results together with the supporting metrics. These included improvements in cross-validated kriging performance when anisotropy was applied (4–40% reduction in CV-RMSE), together with directional consistency metrics such as the wind-alignment index (A), and sector-based rank correlations ρ $(H_{\theta },f_{\theta })$ . By treating anisotropy as an empirically supported feature rather than imposing it solely from the wind direction, this approach integrates spatial statistics with meteorological observations to interpret ΔCH₄ plume structure. Previous studies show that selectively accounting for anisotropy can improve the accuracy and interpretability of CH₄ maps. ,

Variogram parameters also reflect the plume structure. Flights with steady or moderately variable winds corresponding to wu stability conditions produced longer spatial correlation ranges (∼58–150 m), consistent with coherent plume transport under shear-driven turbulence. In contrast, calm or highly variable winds under xu conditions yielded shorter ranges (∼45–133 m) and larger nugget contributions (up to ∼0.85 of the sill), reflecting rapid decorrelation and fine-scale concentration variability characteristic of convectively dominated turbulence (Table S2, which summarizes fitted variogram parameters for each UAV flight). These patterns indicate that spatial dependence weakens as wind directionality declines, linking ΔCH₄ patterns to atmospheric transport processes.

3.3. Methane Hotspot Structure and Its Wind Dependence

Across all UAV flights, local Gi* analysis showed statistically significant methane hotspot clusters; however, their shape, orientation, and spatial coherence depended strongly on the wind regime (Figure ). Under steady or moderately variable winds, all four October 2024 flights tended to show a hotspot share concentrated in sectors broadly aligned with the wind direction, with D ² ranging from 36.0% to 68.8% and |Δφ| ranging from 1.1° to 56.8° (Figure ). UAV-derived wind direction carries a mean absolute error of ∼26°, increasing to 17–65° under weaker winds. Accordingly, |Δφ| should be interpreted as an approximate indicator of directional consistency rather than precise angular agreement and is most reliable under coherent flow. The 29 October, 12:00 p.m. flight (|Δφ| = 1.1°, D ² = 46.1%) is therefore interpreted with reduced confidence due to its large (∼65°) wind direction uncertainty. Despite this overall alignment, variability in angular deviation across flights highlights that hotspot directionality is not uniform even within the same wind regime. These flights were characterized by relatively directional flow (Circular SD 5.4–46.3°) and variable variogram elongation (b/a = 0.30–0.45 for three of the four flights; the 30 October, 11:00 a.m. flight produced an isotropic variogram despite coherent hotspot directionality, as discussed in Section ).

In contrast, calm or highly variable winds, such as the 29 October 2024 (12:00 p.m.) and all August 2025 flights, produced diffuse, weakly oriented hotspot fields. For these flights, the hotspot share was fragmented across nearly all sectors, with no single direction accounting for more than 20–30% of the total. These cases exhibited significantly higher directional wind spread (Circular SD 39.1–73.8°) and generally higher anisotropy ratios, reflecting a more isotropic spatial distribution. On the other hand, the 30 October 2024 (11:00 a.m.) flight showed a more distributed hotspot share than the morning’s steady-wind case.

Under coherent wind conditions, hotspot patterns showed a strong directional structure. Deviance explained by the GLM (D ²) ranged from 36% to 69%, while the OLS model yielded R ² values of 0.38–0.61, and hotspot alignment with the downwind axis was high (|Δϕ| ≤ 13.5°). These results indicate that under steady wind conditions, ΔCH₄ enhancements form plumelike hotspot patterns aligned with the wind direction. Across all flights, high directional coherence (D ² > 36%) occurred when winds exhibited directional persistence, including both weakly unstable and extremely unstable conditions, whereas low coherence (D ² < 12%) occurred only under extremely unstable, highly variable winds. This behavior is consistent with dispersion theory and previous airborne and UAV-based methane studies demonstrating that coherent winds organize concentration fields along a dominant transport axis. ,

Flight-specific examples illustrate this coupling, as shown in Figure . During the 29 October 2024 flights, hotspot clustering remained strong even as wind variability increased. At 12:00 p.m., although classified as calm/variable based on mean wind speed and circular SD, hotspots extended along the E-SE axis with nearly perfect directional agreement (|Δϕ| = 1.1°, D ² = 46.1%, R ² = 0.51). By 2:00 p.m., increased wind variability rotated the plume, but directionality remained robust (D ² = 55.0%, R ² = 0.61, A = 0.49), forming both downwind and crosswind lobes. A similar pattern occurred on 30 October 2024 at 9:30 a.m., where the hotspot axis nearly matched the downwind direction (|Δϕ| = 5°) despite limited sectoral overlap, illustrating that plume orientation and plume co-occurrence represent distinct facets of hotspot-wind coupling. , Because the alignment metric is based on sectoral co-occurrence rather than angular separation, low A values indicate weak directional correspondence but do not distinguish opposite, orthogonal, or dispersed patterns.

It is important to note that the October 2024 and August 2025 campaigns were separated by approximately nine months. During this period, changes in cattle occupancy, manure accumulation rates, and environmental conditions at the facility have the potential to alter the underlying CH₄ source strength, independent of atmospheric transport. Accordingly, differences in hotspot intensity or ΔCH₄ magnitude between campaigns are interpreted cautiously, with emphasis placed on wind-driven differences in plume structure rather than absolute enhancement magnitude. In contrast, ΔCH₄ hotspots during the August 2025 flights were weaker, more dispersed, and frequently misaligned with the mean wind direction, reflecting low wind speeds and high directional variability that limited plume coherence and supported isotropic or multidirectional dispersion. GLM deviance explained dropped to 1.6–11% and fitted hotspot directions deviated by more than 50°. However, moderate values of the alignment index (A = 0.22–0.48) and localized correlations (ρ up to +0.61) indicate that some short-range directional tendencies persisted, likely reflecting intermittent plume fragments under variable transport. Such degraded coherence under variable winds is consistent with airborne flux studies showing that inconsistent advection disperses CH₄ laterally and reduces detectability of dominant transport pathways. , Reduced directional coherence has also been reported under stable stratification but through a different mechanism. Gucci et al. (2025) attribute this behavior to one-component turbulence associated with suppressed mixing in stable conditions, whereas the present xu flights occurred under strongly unstable convective turbulence that promotes intense mixing and plume fragmentation.

Across all flights, UAV-derived ΔCH₄ fields were most reliably interpretable when the winds exhibited at least moderate directional coherence. In practice, plume alignment and hotspot patterns were consistently recovered when mean wind speeds exceeded roughly 2 m s^–1 and turbulence intensities were below approximately 0.35. Under these conditions, GLM directionality (D ²) exceeded 35%, anisotropy ratios (b/a) remained below 0.5, and kriged surfaces reproduced elongated, source-anchored plumes originating from the occupied pens. Conversely, flights with weaker winds (U̅ < 2 m s^–1) or higher turbulence (TI > 0.40) produced diffuse or isotropic hotspot patterns, directional misalignments exceeding 50–100°, and substantially reduced spatial correlation-conditions under which source attribution became far less reliable. These empirically derived ranges should be interpreted as contextual benchmarks rather than universal thresholds, highlighting conditions under which spatial pattern confidence is enhanced or degraded.

These results suggest that CH₄ hotspot geometry appears more strongly associated with advective coherence rather than simply the presence of elevated concentrations. Steady winds impose a coherent plume axis that results in well-defined spatial organization, whereas calm or variable winds fragment plume structure and produce multioriented patterns that increase spatial complexity and decrease mapping confidence. Although a positive association was observed between mean wind speed and directionality metrics (e.g., ρ = 0.4 for wind speed versus D ², R ², and |ρ­(H_θ ,f_θ )|), the relationship was not statistically significant (p > 0.3), likely reflecting the limited sample size (n = 7 flights). However, across the individual flight cases, multiple directional diagnostics, including plume alignment (Δφ), sectoral overlap (A), and variogram anisotropy, consistently indicated stronger directional organization under more coherent wind conditions, reinforcing the role of advection in shaping the plume geometry across flow regimes. Combining complementary metrics, angular alignment (GLM/OLS), sectoral overlap (alignment index, A), and rank correlation (ρ), provides cross-validating insights while reducing reliance on any single metric. , This multimetric approach strengthens confidence that observed ΔCH₄ hotspots represent true emission-driven patterns modulated by wind, rather than artifacts of UAV sampling or spatial interpolation.

3.4. Methane Enhancement Patterns across Feedlot Features

To evaluate how plume geometry and directionality corresponded spatially with feedlot features, anisotropy-corrected ΔCH₄ mosaics and zonal statistics for occupied pens and lagoons were examined. Across all flights, kriged maps generally showed higher mean ΔCH₄ over occupied pens, though enhancement magnitude varied substantially (0.06–0.54 ppm; Table ), and spatial patterns depended strongly on the wind regime (Figure ). The number of occupied and empty pens varied slightly among flights because pen occupancy was determined using contemporaneous feedlot inventory records collected on each sampling date. Although individual occupied pens exhibited internal variability in the ΔCH₄ intensity, the aggregated zonal mean for occupied pens was consistently higher than that for empty pens across all flights with both pen types present. On 29 October, the mean ΔCH₄ was 0.42 ppm for occupied pens (n = 109 pens) compared with 0.30 ppm for empty pens (n = 5 pens). Similarly, on 22 August, occupied pens showed a mean of 0.35 ppm (n = 103 pens) versus 0.32 ppm for empty pens (n = 11 pens). On 30 October, all monitored pens were occupied (n = 114 pens), resulting in a lower mean signal (0.10 ppm), although the enhancement remained spatially associated with the production area. It is important to note that because measurements were taken at 20 m AGL, these “hotspots” represent advected plumes rather than stationary surface sources; therefore, a low signal within an occupied pen boundary often reflects plume detachment or horizontal displacement by wind rather than a lack of methane production. During the October 2024 flights, ΔCH₄ observations and interpolated surfaces exhibited relatively coherent spatial organization consistent with downwind-oriented methane enhancement patterns originating from the central pens (Figure ; Figure S2), a morphology characteristic of advection-dominated dispersion. In contrast, the August 2025 flights produced more spatially fragmented and weakly organized enhancement patterns, reflecting more isotropic dispersion under calm and variable wind conditions despite comparable peak ΔCH₄ values (0.85–1.28 ppm; Table ). These contrasts indicate that wind conditions modulate the spatial expression of ΔCH₄ enhancements, influencing how the underlying signals are formed and observed.

4. Zonal ΔCH₄ (ppm) Statistics by Structural Class (Pens and Lagoons) for Each UAV Flight, Showing Mean, 95th-Percentile, and Maximum Values.

UAV Flight (Date and Time)	Pens Mean	Pens P95	Pens Max	Lagoons Mean	Lagoons P95	Lagoons Max
29 Oct 2024 12 p.m.	0.54	0.62	1.34	0.18	0.22	0.69
29 Oct 2024 2 p.m.	0.28	0.32	0.80	0.11	0.14	0.32
30 Oct 2024 9:30 a.m.	0.12	0.16	0.87	0.18	0.26	0.95
30 Oct 2024 11 a.m.	0.06	0.09	0.60	0.03	0.04	0.16
22 Aug 2025 11:45 a.m.	0.33	0.45	1.28	0.27	0.38	0.68
22 Aug 2025 12:30 p.m.	0.36	0.42	0.85	0.30	0.34	0.47
22 Aug 2025 2 p.m.	0.35	0.46	1.22	0.22	0.26	0.53

Though the wind-regime contrasts explain much of the observed plume geometry, interpreting where enhancements “belong” on the ground also depends on how sources are arranged within the facility. In this feedlot, manure accumulation occurred within occupied pens rather than in discrete piles, resulting in colocated enteric and manure emissions at the pen scale. Once plumes mix, UAV-based concentration measurements cannot separate these components; accordingly, occupied pens are treated as composite emission zones. The zonal analysis therefore focuses on distinguishing pen-associated enhancements from ΔCH₄ observed over lagoon surfaces, which primarily represent areas that plumes may traverse rather than independent emission zones, thereby helping to identify spatially consistent emission zones versus areas dominated by advective plume transport.

During the October flights, conducted under weakly unstable (wu) conditions when the plume structure was most coherent, ΔCH₄ maxima were spatially contiguous with upwind pen areas, consistent with advection from nearby sources, although concentration mapping alone cannot confirm source origin. The highest 95th-percentile values and maximum ΔCH₄ enhancements (up to 1.34 ppm) occurred inside occupied pen boundaries (Table ). Specifically, the 29 October, 12:00 p.m. flight recorded the study-high maximum of 1.34 ppm, while the 2:00 p.m. flight reached 0.80 ppm as wind speeds increased. While the low-wind conditions at 12:00 p.m. resulted in a more diffuse accumulation of ΔCH₄, the spatial mosaics for the 2:00 p.m. flight and the 30 October (9:30 a.m.) flight show clearly structured downwind elongation extending from the active pens (Figure , Figure S2). During the 30 October 9:30 a.m. and 11:00 a.m. flights, enhancements remained pen-centered and downwind-aligned, with maxima of 0.87 and 0.60 ppm, respectively (Table ). These patterns are consistent with advection-dominated dispersion, where coherent winds tended to preserve plume structure over facility-scale distances.

In contrast, ΔCH₄ patterns during the August 2025 flights reflected the breakdown of plume coherence under calm, variable winds, and extremely unstable (xu) conditions.

Although pen means (0.33–0.36 ppm) and maxima (0.85–1.28 ppm) remained elevated, spatial patterns were diffuse, weakly oriented, and shifted between flights. The lower pen-mean ΔCH₄ values observed on 30 October (0.06–0.12 ppm) relative to the August flights (0.33–0.36 ppm) are also consistent with the shallower convective boundary layer expected under weakly unstable (wu) conditions, where shear-driven turbulence confines mixing to a relatively shallow layer, potentially reducing the vertical dilution of near-surface emissions compared with the deeply mixed convective boundary layer characteristic of the extremely unstable (xu) August flights.

Directionality metrics corroborated this reduction (D ² ≤ 0.11; R ² ≤ 0.11), and no stable plume axis emerged (as discussed in Section ). These disorganized patterns are consistent with isotropic or multidirectional dispersion expected under variable winds. ,

Across most flights, lagoons displayed low-level and intermittent ΔCH₄ anomalies (mean ≤ 0.30 ppm; maxima ≤ 0.95 ppm). While liquid manure storage systems associated with dairy or swine operations can be substantial methane sources, the lagoons at this feedlot function as runoff storage, with less persistent anaerobic conditions and infrequent inputs due to low rainfall. Similarly, previous measurements at a comparable commercial feedlot in Australia showed that runoff ponds contributed less than 1–2% of pen and manure stockpile emissions. Accordingly, elevated CH₄ concentrations observed over lagoon areas are most plausibly attributed to advected plumes from nearby occupied pens rather than local lagoon emissions. Because UAV-based concentration mapping at a fixed altitude cannot distinguish between locally emitted CH₄ and CH₄ transported from upwind sources, an independent lagoon emission signal cannot be resolved from these measurements. However, differences in ΔCH₄ between lagoons and pens may also reflect near-field plume dynamics. Surface-level lagoon emissions over smooth terrain may remain confined within a shallow layer, whereas pen emissions originate from elevated, obstacle-rich environments that enhance turbulence and vertical mixing. Such effects of source height and surface structure on plume behavior indicate that observed differences likely reflect both emission strength and atmospheric transport. An exception occurred during the 30 October 2024 (9:30 a.m.) flight, when mean ΔCH₄ values over lagoon areas slightly exceeded those over the pens (Table ). Under the steady southwesterly winds of this flight (mean direction of 211°), the southern lagoons were positioned upwind of the pen area, and the plume advected northeastward across the lagoon polygons before reaching the occupied pens. This geometry likely caused the kriged concentration field to register higher values within lagoon boundaries relative to pens rather than reflecting elevated local lagoon emissions. Because zonal statistics represent concentration intensity computed from uniform-resolution kriged surfaces rather than area-integrated quantities, this result is consistent with advective transport from the pen source area. Similar transport-driven patterns have been reported in receptor-based analyses such as conditional bivariate probability function (CBPF) studies at other livestock facilities. Empty pens exhibited intermediate ΔCH₄ values, reflecting partial plume transport from nearby occupied pens.

To assess the signal contribution of livestock, we compared mean ΔCH₄ concentrations between occupied and empty pens for all flights, where both pen types were present simultaneously. During the 29 October 2024 (12 p.m.) flight, occupied pens showed a mean enhancement of 0.561 ppm, compared with 0.349 ppm for empty pens; during the 2 p.m. flight, values were 0.281 and 0.243 ppm, respectively. Across the August 2025 flights, occupied pen means ranged from 0.336 to 0.377 ppm, while empty pens averaged 0.231–0.383 ppm. To quantify the statistical reliability of these zonal means, we calculated the standard error (SE) for each structural class, which accounts for intraclass spatial variability across the individual polygons (detailed in Table S3). The SE for occupied pens typically remains low (0.01–0.02 ppm), confirming that the observed ΔCH₄ patterns are spatially robust. In contrast, higher SE values within lagoon zones reflect both smaller sample sizes and the transient nature of plumes advecting over those areas rather than the instrument noise. These results indicate that occupied pens generally exhibited higher or comparable mean enhancements, while occasional convergence between pen types likely reflects increased plume mixing under variable wind conditions. Altogether, spatial mosaics, directional metrics, and zonal statistics show that occupied pens frequently coincide with elevated ΔCH₄ signals in this facility. However, plume displacement under advective conditions may offset concentration maxima from underlying sources, underscoring the strong dependence of ΔCH₄ spatial patterns and interpretability on the wind regime. UAV-based concentration mapping at 20 m AGL therefore reflects the combined effects of emission strength, atmospheric transport, and turbulent mixing. Consequently, observed enhancement patterns cannot be interpreted as direct measurements of emission flux from individual surface features. Robust source attribution would require coupling concentration observations with atmospheric transport modeling or inverse dispersion approaches.

3.5. Implications for Feedlot Methane Management and Broader Methane Monitoring Systems

The results of this study provide practical insights for interpreting facility-scale methane measurements in commercial feedlots and other heterogeneous emission environments. Across all campaigns, occupied pens frequently coincided with elevated ΔCH₄ signals when sampled under favorable wind conditions. The consistency of these signals across seasons and wind regimes suggests that pen-associated sources consistently contribute to near-field methane concentration patterns when sampled under stable background methane conditions and adequate wind transport. In contrast, atmospheric transport, driven largely by wind conditions, controls the detectability, spatial distribution, and interpretation of the observed ΔCH₄ fields. Under steady or moderately coherent winds, ΔCH₄ plumes exhibited clear downwind alignment and retained spatial structure relative to upwind pen areas, enabling a higher confidence in spatial attribution at the feature scale. On the other hand, under calm or highly variable winds, plume coherence deteriorated and ΔCH₄ fields approached isotropy, substantially reducing attribution confidence despite the continued presence of elevated concentrations. This study's results support the potential usefulness of applying flow regime screening as an interpretive filter in UAV-based methane surveys, whereby measurements collected under modest directional coherence (U̅ ≳ 2 m s^–1; TI ≲ 0.35) are more likely to yield spatially interpretable concentration fields, although data collected under weak or turbulent winds should be interpreted with caution. Where a sonic anemometer or nearby weather station is available, atmospheric stability class provides a complementary screening criterion, in this study, wu conditions (ζ between −0.04 and −0.10) consistently supported coherent, spatially interpretable plume structures, while xu conditions (ζ < −1) were associated with fragmented, near-isotropic ΔCH₄ fields regardless of instantaneous wind speed, suggesting that ζ could serve as a practical preflight or posthoc quality filter in future UAV-based surveys. Such interpretive selection is especially important as measurement-informed CH₄ reporting and verification frameworks increasingly combine facility-scale surveys with broader atmospheric constraints.

Although feedlot configurations differ in manure handling and facility layout, the wind-informed spatial framework developed here is broadly transferable to other methane-emitting systems characterized by colocated sources and complex near-surface flow. The framework applies whether manure is stored within pens, in centralized piles, or in adjacent infrastructure because it is based on diagnosing plume patterns and spatial observability rather than assuming source behavior. At the same time, the results highlight a limitation of UAV-based concentration measurements, such as once plumes mix, enteric and manure emissions originating within the same spatial unit cannot be reliably disaggregated. While multispecies measurements with distinct emission ratios (e.g., CH₄–NH₃ or CH₄–CO₂) could potentially aid source discrimination, rapid near-field mixing within occupied pens may limit the effectiveness of such approaches under coherent advection. The spatial consistency of pen-associated ΔCH₄ signals across contrasting wind regimes further suggests that combining high-resolution concentration mapping with wind information and facility layout could support feature-scale CH₄ flux estimation in future studies.

A further practical implication is the ability to distinguish spatial features that are consistently associated with methane enhancement from surfaces that are intermittently intersected by transported plumes. In this study, ΔCH₄ anomalies over lagoon areas may arise from either advected pen plumes or local lagoon emissions; concentration mapping alone cannot distinguish between these mechanisms. Such behavior is consistent with an advective plume overpass rather than sustained local emissions. Wind-informed analysis, including anisotropy metrics, plume-wind alignment, and zonal statistics, proved essential for identifying these distinctions and avoiding overinterpretation of transient concentration enhancements.

Although this study focuses on a single feedlot facility and a limited number of UAV flights (n = 7), the results highlight methodological considerations relevant to interpreting high-resolution methane concentration measurements in heterogeneous emission environments. In particular, the analysis demonstrates the importance of accounting for wind-driven transport in interpreting spatial patterns in ΔCH₄ fields. Previous studies have also shown that, at larger scales, methane source partitioning can remain prior-dependent even when atmospheric constraints are applied, highlighting persistent challenges in methane attribution across monitoring frameworks. Parallel efforts to develop high-resolution, sector-resolved methane inventories for local-scale modeling further underscore the importance of spatially explicit methane information for policy-relevant analyses. In this context, integrating anisotropy-aware geostatistical analysis with wind-informed interpretation provides a practical framework for assessing when UAV-derived ΔCH₄ concentration fields retain meaningful spatial structure and when attribution uncertainty is likely to increase. These findings are based on observations from a single commercial feedlot across seven UAV flights spanning two seasonal campaigns. While the relationships identified between wind conditions and ΔCH₄ spatial structure are physically interpretable, the limited sample size and site-specific characteristics may constrain direct generalization to other facility types or climatic conditions. Further studies across diverse facilities and environmental conditions would help evaluate the broader applicability of these findings.

3.6. Outlook on Improvements to the Measurement Method

This study demonstrates that UAV-based measurements can resolve fine-scale, wind-dependent patterns of ΔCH₄ fields in heterogeneous emission environments. Several refinements could further improve the accuracy and interpretability, particularly under conditions that challenge spatial attribution. Expanding measurements across a broader range of meteorological conditions and seasons would reduce sampling uncertainty and better capture plume variability. Integrating a triaxial ultrasonic anemometer on the UAV would enable direct measurement of turbulence and small-scale wind variability, improving interpretation under weak or rapidly changing winds. Until a co-mounted anemometer is available, UAV-derived wind estimates should be interpreted as flight-average approximations subject to the systematic biases documented in the validation results. Incorporating atmospheric stability screening (e.g., ζ = z/L), where surface-layer flux data are available, would provide a practical preflight or post hoc quality filter to identify conditions most likely to yield spatially interpretable ΔCH₄ fields. Multialtitude flight transects would further resolve vertical ΔCH₄ structure, plume depth, and background variability. Together, these advances would reduce transport-related uncertainty and strengthen the confidence in the spatial interpretation of high-resolution UAV-based CH₄ observations. While the concentration-based approach resolves spatial patterns, it does not directly quantify emissions from individual source types. Coupling high-resolution concentration mapping with inverse dispersion modeling or mass balance approaches in future campaigns would enable quantification of source-specific emission contributions from pens and lagoons. These measurement strategies would also support more robust estimation of facility-scale methane fluxes in future UAV studies.

The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acs.est.5c18173.

• Description of methane detection sensor; variogram modeling and anisotropy detection methods (including empirical variogram computation, model fitting, cross-validation, and anisotropy testing); software environment; Table S1 (instrument specifications); Table S2 (fitted variogram parameters); Table S3 (zonal ΔCH₄ statistics); Figure S1 (feedlot pen layout); Figure S2 (spatial distribution of methane enhancements) (PDF)

The authors declare no competing financial interest.