AVIRIS-4 is a state-of-the-art imaging spectrometer with identical core components as NASA JPL's AVIRIS-3 and the EMIT spectrometer . The spectrometer is equipped with a 1280-pixel sensor array and records hyperspectral data in 328 bands spanning the ultraviolet (UV) to the shortwave infrared (SWIR). In practice, 1241 pixels receive sufficient illumination and SNR, and 287 bands are retained for data processing. Detailed sensor specifications are provided in . Compared to its predecessor it offers enhanced stability, spatial sampling interval (hereafter referred to as spatial resolution) and signal-to-noise ratio (SNR) (see Table ).

The data for this analysis was acquired during a single-blind controlled release experiment organised by the Environmental Assessment and Optimization Group at Stanford University between the 16 and 20 September 2024 at the TotalEnergies Anomalies Detection Initiatives (TADI) site in Lacq in the south of France (latitude: 43.412°, longitude: -0.636°, elevation a.m.s.l.: 95 m) (see Fig. a). A total of 13 commercial and academic teams, using a range of technologies – including continuous monitoring, vehicle-based measurements, drones, airborne in-situ measurements, remote sensing from aircraft, and satellites – participated in the experiment. The results of all teams were collected and analysed in . On each campaign day (08:00–18:00 CEST), up to 9 individual controlled releases with rates varying between 0.02 and 350 kgCH4h-1 were conducted at different unknown heights between 0.01 to 6.5 m above ground and at different unknown locations on the study site (see Fig. b). Each release lasted for 45 min and was followed by a 15 min break before the start of the next release. In some periods, no CH₄ was released to enable the detection of false positives. Additionally, the wind speed was measured using a ZX 300 Doppler wind lidar positioned 100 m from the emission sources. The instrument recorded horizontal and vertical wind speeds, as well as wind direction, at preselected heights between 10 and 300 m above ground level, with a temporal resolution of approximately 20 s. Participating teams were aware of the timing of releases while locations and flow rates of the releases as well as the wind data were only made available after all teams had submitted their initial emission estimates. Details of the release experiment, the participating teams and the synthesis can be found in . For the campaign, AVIRIS-4 was mounted on a hydraulic stabilisation mount and built into a Cessna 208B Grand Caravan EX. The aircraft flew over the release site in either north-south or east-west direction at different altitudes of 12 000, 9000, 6000, 4200 and 3300 ft or 3660, 2740, 1830, 1280, 1000 m above mean sea level (a.m.s.l.) (see Fig. a). This resulted in average spatial resolutions of 2.0, 1.5, 1.0, 0.7 and 0.5 m across-track and 0.35 m along-track. For the remainder of the article, all wind speed heights are given in metres above ground level and all flight altitudes in feet a.m.s.l.

We retrieved CH₄ maps from the AVIRIS-4 radiance cubes using the computationally efficient matched filter approach following and further refined by . The filter detects a known signal within a noisy background by enhancing the signal relative to the noise, effectively maximising the output signal-to-noise ratio under the assumption of additive Gaussian noise.

To estimate the CH₄ emissions, we used the integrated mass enhancement (IME) and cross-sectional flux (CSF) method implemented in the Python library for data-driven emission quantification (ddeq) . We used the CSF for longer plumes and more turbulent conditions as it averages the fluxes along several cross-sections. Conversely, the IME was used for short plumes and plumes that deviate from a Gaussian plume shape such as for overlapping double plumes. Both methods assume steady-state conditions of wind speed and emission rate. Limits of this assumption are further discussed in Sect. .

All mass-balance based methods require an estimate of the wind speed U. Ideally, U would correspond to the effective wind speed Ueff, which is the mean speed at which the plume is transported . However, as the vertical CH₄ profile is unknown, we used four different approaches to obtain a wind speed estimate:

10 m wind speeds U10 from ERA5 reanalysis data as used for the initial reporting in as ground-based lidar measurements were not available prior to unblinding.

We also conducted a sensitivity analysis using lidar wind speeds at other elevations above ground level.

Below we describe how uncertainty components are estimated and propagated for each input to the emission quantification.

In contrast to previous efforts, this new generation of controlled release experiments was planned to reflect more realistic natural conditions. While this allows to assess sensor performance in diverse terrain and meteorological conditions it also introduces limitations associated to different surface coverage, cast shadows and cloud conditions (see Fig.  for detailed meteorological setting during all experiments, and Figs.  and  as well as Tables  and  in the Appendix for wind information). Despite these challenges, we were able to fly 100 overpasses at different hours of the day (see Fig. a) and at five altitudes (see Fig. b), which allowed us to evaluate the influence of wind speeds and spatial resolution on the CH₄ detection and emission estimation. Flights at all flight levels were only scheduled for the first release in the morning and afternoon after refuelling. The atmospheric stability was estimated to be neutral to unstable for all observations based on the Pasquill stability classes using U10.

Figure (upper row) presents three optimal examples of plumes resulting from three different releases. The plumes appear largely linear, with minimal influence from turbulence, which is favourable for emission estimation. For stronger sources, retrieval noise is barely noticeable, but at lower intensities – such as the 26.4 kgCH4h-1 release – it can interfere with the plume signal and hinder accurate attribution of enhanced pixels (see Sect. ).

The lower row in Fig.  shows three turbulent plumes observed during overpasses at 4200 ft, where local enhancements caused by turbulent eddies are clearly visible. In these cases, the CSF method outperforms the IME approach, as the effect of turbulence is reduced through averaging across multiple cross-sections.

In addition to challenging conditions, there was also a case where turbulence impeded emission estimation, shown in Fig. . A change in wind direction prior to the overpass appears to have caused a large, dispersed “blob” of CH₄ enhancements. Since these conditions violate the steady-state assumption, this case was excluded from emission estimation.

The median noise level of CH₄ maps was estimated to be around 450 ppm m or 0.3 gCH4h-1 for the data of the controlled release experiment. Out of 100 overpasses, plumes were detected on 68 instances (Fig. ). In the most favourable case, the smallest observed plume corresponded to a 1.45 kgCH4h-1 release at U10=0.76 m s⁻¹ and a flight altitude of 4200 ft, representing the best-case detection limit for AVIRIS-4. Under typical conditions, plumes from releases of 5.5 kgCH4h-1 and above were consistently detected at altitudes ≤4200 ft, with the exception of two overpasses where shadows from surface infrastructure obscured the signal. At higher flight altitudes (6000–12 000 ft), detection performance was more constrained: for release rates ≤9.23 kgCH4h-1, only one plume was detected (6000 ft, U10=1.3 m s⁻¹), while the others could not be observed due to the combined effect of higher winds and lower emissions. The original objective of conducting observations at multiple flight altitudes was to determine an altitude-dependent detection limit. However, because the CH₄ release rates were not known in advance, the largest release event captured at all five altitudes was metered at only 9.23 kgCH4h-1. This emission rate was below the detection threshold at altitudes above 6000 ft and therefore remained undetectable in those overpasses. For the overpasses below 6000 ft, we computed the probability of detection (PoD) for AVIRIS-4 according to as a function of reported emissions Qrep, U10 and flight altitude h̃ using the flags “detected” and “not detected” by optimising the predictor and inverse link functions. This resulted in the following PoD function which is plotted in Fig. . 21PoD=1-1+1.03×10105.18×108Q1.93h̃10003.88U10+97.09.97-1.84

We were able to estimate the emission from 67 of the 68 detected plumes, 54 of which were estimated using the CSF method and 13 using the IME method. Figure  shows the reported versus estimated CH₄ emissions using four different wind speed inputs. As outlined in Sect. , the initial CH₄ emission estimates were calculated using ERA5 U10, shown in subplot (a) of Fig. . This approach yields a relatively weak correlation, with a fitted slope of only 0.53 and an R2 value of 0.55. Replacing ERA5 data with lidar-measured U10 in subplot (b) of Fig.  substantially improves the agreement, increasing the slope to 0.65 and an R2 value of 0.73. This highlights the limitations of reanalysis wind data for accurate emission quantification (further shown in Fig. ). As a result, the use of ERA5 introduces both correlated and uncorrelated uncertainties in emission estimates that are difficult to quantify or correct.

Even when using on-site lidar wind speeds (Fig. b), biases remain: emission rates for small release events tend to be overestimated, while large releases (e.g. at 80.1 and 290 kgCH4h-1) are significantly underestimated. This behaviour can be explained by plume dynamics: Small release events result in short plumes which remain near the emission height (<10 m for all releases), making the use of U10 prone to overestimation. In contrast, large releases produce longer plumes that undergo greater vertical mixing. The actual effective transport height may thus be above 10 m, resulting in an underestimation of emissions when using U10. Additional influencing factors are specific to the release equipment, such as the outlet ejection velocity and whether the emission was oriented horizontally or vertically.

These limitations highlight the importance of estimating an effective wind speed (Ueff) that accounts for both source height and vertical mixing. Subplots (c) and (d) in Fig.  compare two approaches: the method of , which accounts only for vertical mixing, and the method developed in this study, which accounts only for source height. In subplot (c), the overall fitted trend lies close to the 1:1 line, but the estimates for small releases are substantially worse than when using U10. This reflects the fact that derived the linear relationship between U10 and Ueff for GHGSat, which has a coarser spatial resolution (50 × 50 m). At that scale, plumes have more time to mix vertically and are therefore transported by winds stronger than U10. In contrast, subplot (d) shows a poorer overall trend than (c) due to the strong influence of large release events, but the estimates for small releases improve considerably. This suggests that short plumes are well captured because they remain close to the emission height, whereas vertical mixing is insufficiently accounted for in the case of larger releases.

To further investigate this hypothesis of strong vertical mixing, we incorporated lidar wind speeds at 20 and 38 m and calculated the uncertainty-weighted root mean squared error (RMSE) and relative mean bias error (MBE) between estimated and reported CH₄ emissions, as shown in Fig. . The results confirm that using Usrc substantially improves emission estimates for low intensity release events. For release events above 30 kgCH4h-1, however, using Usrc tends to underestimate emissions and performs worse than estimates based on U10, U20 and U38. Although the relative MBE decreases for larger releases, the high relative RMSE indicates substantial variability around the true values. This pattern may reflect the greater influence of turbulence on longer plumes compared to shorter ones.

In addition to source strength and therefore plume length, absolute wind speed appears to significantly influence the accuracy of emission estimates. This is illustrated in Fig. , which shows the scaling factor required to align estimated emissions with reported values as a function of (a) plume length and (b) effective wind speed. While subplot (a) of Fig.  supports the previously discussed hypothesis regarding plume length, subplot (b) reveals that lower wind speeds are associated with larger and more variable scaling factors. This observation aligns with the findings of , who reported reduced accuracy in emission estimates across various techniques under low wind speed conditions. This is likely due to the increased variability typically observed at lower wind speeds. In contrast, we did not observe larger scaling factors for larger coefficients of variation (CoV) in wind direction in subplot (c) of Fig.  as discussed in . The reason for this is that our method does not depend on wind direction, as we do a nearly instantaneous measurement. The large spread in angles between the wind direction and the curve fitted to the plume in subplot (d) further highlights the strong influence of wind turbulence on the observed plumes.

This hypothesis is further supported by individual cases where estimated emissions diverge from reported values, as illustrated in Fig. . Subplot (d) shows that the CH₄ fluxes across different cross-sections fluctuate strongly between 100 and 200 kgCH4h-1 due to turbulent wind, likely reflecting both temporal variability in wind speed and changes in plume height that exposed it to different wind regimes. In such cases, one might consider using only CH₄ enhancements close to the source, such as those from the first cross-section, where Ueff is expected to better approximate the wind speed at source height. However, this example shows that even this approach leads to underestimation, indicating that the measured wind speeds do not accurately reflect actual wind conditions. An analysis of the wind speed during the two minutes of the overpass reveals that U10 varies between 1 and 3 m s⁻¹. For comparison, a 20 m plume under a 1 m s⁻¹ wind has a residence time of about 20 s, which matches the sampling interval of the wind lidar. As a result, the wind speed fluctuations visible in the plume cannot be resolved by the lidar, and the underestimation can likely be attributed to larger-than-expected temporal variability that is not captured at the instrument's temporal resolution.

The analysis of uncertainty contributions to total emission estimate uncertainty (Fig. ) indicates that wind speed is the dominant factor for both the CSF and IME methods. Most of this contribution arises from the natural variability of wind speed, with additional influence from uncertainty in the effective wind speed. In comparison, measurement errors in wind speed account for only a minor portion of the overall uncertainty.

Lastly, one source of deviation between estimated and reported CH₄ emissions is the presence of cloud shadows over the release site as shown in Sect. , leading to the strong underestimations of the 80.1 kgCH4h-1 release event observed in Fig. . Despite this underestimation, the plumes were still reliably detected, indicating that observations under suboptimal cloud conditions can still be valuable e.g. for leak detection.

The high spatial resolution of AVIRIS-4 offers the unique opportunity to estimate the height h of an emission source based on the length of the shadow ls in the RGB image (Fig. a) cast by the emission source using trigonometry: 22h=lstan⁡(SZA) Alternatively, the height can be estimated in the same way from the horizontal separation of the starting points of the two plumes (Fig. b). With increasing distance, the two plumes move together more closely, suggesting that the plume is pushed towards the surface directly after the release. Knowledge of the emission height is an important parameter for emission estimation, as it can be used to determine the effective wind speed, which is a critical input for estimation estimation. In the example shown in Fig.  with an SZA of 50° and a spatial resolution of 0.53 m, the emission plume at the stack must be 6.4,± 1.1 m above ground which is in agreement with the true emission height of 6.5 m.

We assume that this technique can be reliably applied only if the measured shadow length exceeds its measurement uncertainty by a sufficient margin. The uncertainty in the shadow length is dominated by pixel discretization at the shadow boundaries, where at most one mixed pixel can occur at both the upper and lower edge of the shadow. Requiring the shadow length to be at least twice this uncertainty ensures that the shadow is sufficiently resolved. Under this criterion, the minimum emission height that can be resolved is given by 23h>2⋅lptan⁡(SZA) For the campaign discussed in this study, this minimum height is shown in Fig. .

Following the success of the AVIRIS Classic and AVIRIS-NG sensors in detecting and quantifying CH₄ emissions as demonstrated in numerous previous studies, this study explores the potential of their successor, AVIRIS-4. Although AVIRIS-4 was primarily developed for surface and vegetation studies, our results show that CH₄ columns can be determined with an unprecedentedly high spatial resolution, enabling the detection of short plumes from low intensity sources. In combination with the enhanced SNR, the detection limit is reduced to 5.5 kgCH4h-1 under good weather conditions and down to below 1.5 kgCH4h-1 under ideal conditions. Because the campaign took place in mid-September, we expect the detection limit could be further reduced under more favourable illumination conditions. As discussed in Sect. , we were not able to determine an altitude-dependent detection limit for AVIRIS-4, which complicates direct comparisons with other airborne sensors. For instance, studies with AVIRIS-NG operated at altitudes between 3000 and 6000 m report detection limits of 10–16 kgCH4h-1 under favourable wind conditions e.g.. In our case, the lowest release of 9.23 kgCH4h-1 could not be detected at a comparable altitude of 2740 m, likely due to higher wind speeds. This makes it difficult to assess whether and by how much the detection limit has improved. In , the lowest detected release was 2.3 kgCH4h-1, but at a much lower flight altitude of 430 m and under higher wind speeds of 3–5 m s⁻¹. report a detection limit of 15 kgCH4h-1 at a flight altitude of 6000 m with wind speeds of 0.5 m s⁻¹. Overall, comparing detection limits across studies is challenging, as they depend strongly on flight altitude, wind speed, and spectral albedo .

Decreasing the detection limit is pivotal because low intensity CH₄ sources are more numerous than high-emitting ones e.g.,. Consequently, accurate estimates of total CH₄ emissions depend on detecting smaller sources. For instance, based on the best detection limit of AVIRIS-NG of 15 kgCH4h-1 reported in and the distribution of oil production sites in Romania across the outlined scenarios, AVIRIS-NG was able to detect between 45 % and 62 % of total emissions. In contrast, assuming a detection limit of 5.5 kgCH4h-1, AVIRIS-4 would increase this detection coverage to approximately 67 %–81 %. Moreover, the detection limit of 5.5 kgCH4h-1 achieved by AVIRIS-4 effectively enables the identification of all point sources listed in the E-PRTR registry, which mandates reporting for emissions exceeding 100 000 kgCH4yr-1 (equivalent to 11.4 kgCH4h-1) .

This study, along with comparisons to other airborne imaging spectrometers with higher spectral but lower spatial resolution such as MAMAP2DL e.g., demonstrates that the trade-off of higher spatial and slightly lower spectral resolution is beneficial for detecting small-scale CH₄ enhancements from low intensity sources, whose plumes typically extend only a few decimetres to a few metres.

The noise level of the CH₄ maps was estimated as the standard deviation of the retrieved columns over the brightest 50 % of pixels. This resulted in a noise level of AVIRIS-4 of 450 ppm m at an average resolution of 0.5 m which is comparable to reported values of AVIRIS-NG at 5 m resolution for suboptimal illumination conditions e.g.. Such noise levels are expected, given that the campaign was conducted in mid-September under low solar zenith angles (SZAs). In addition, negative values were not masked during the CH₄ retrieval, which increases the apparent noise.

The current study also shows that owing to the higher SNR and higher spatial resolution, emissions can also be detected and estimated with less illumination and under suboptimal surface and atmospheric conditions, which are characterised by inhomogeneous albedo, strong turbulence, cast shadows and cloud shadows (see Sect. ), compared to previous controlled release experiments with AVIRIS-NG e.g.. For example, the higher spatial resolution allows for a more accurate filtering for shadow pixels and albedo artefacts which, if undetected, could lead to biases in emission estimates, as outlined in Sect. . This capability allows AVIRIS-4 to be effectively applied to built-up sites with heterogeneous surface albedo and cast shadows, conditions commonly encountered around CH₄ sources in the oil, gas, and coal mining sectors.

However, the higher spatial resolution also results in new challenges. One of them is the occurrence of double plumes originating from plume shadows illustrated in Fig. . We corrected for this artefact when the true plume and its shadow were clearly separated, but its impact on retrieved CH₄ enhancements requires further analysis. In this context, a recent study by systematically investigated the effect of different observation and illumination geometries on the retrieved CH₄ maps (i.e. parallax effect) and the resulting emission estimates. They showed that large VZAs and SZAs can lead to artificial elongation or compression of plumes along the plume direction. This bias in apparent plume length L directly propagates into emission estimates and likely also occurred in the observations analysed in this study. However, their influence is probably masked by the comparatively large variability in wind speed. Furthermore, found that the parallax effect substantially reduces the PoD due to lower apparent CH₄ enhancements. In their simulations, the PoD varied between approximately 0.5 and 0.8 depending on the angular configuration. For the present study, the influence of parallax effects is likely minor, as the detection outcomes shown in Fig.  are primarily controlled by wind speed and flight altitude. The few non-detected plumes with emission rates exceeding 5 kgCH4h-1 at low wind speeds are instead attributable to overlaps with retrieval artefacts. also demonstrated that when the effective wind speed Ueff is calibrated against the 10 m wind speed U10 using L, biases in L translate into systematic errors in the calibration itself. As a consequence, emission estimates exhibit errors below 10 % for mid-latitude summer conditions, but can reach up to 30 % for wintertime observations. In the context of this study, the parallax-induced bias in Ueff is only relevant for emission estimates derived using the Ueff parametrisation of and does not affect estimates based on wind speeds at the source height. To mitigate the effect of viewing geometry, recommended to explicitly account for observation and illumination geometry in the planning of flight paths for airborne sensors and to calibrate Ueff using plume simulations that match the angular configuration (“train as you measure”). Overall, additional work is needed to correct for the parallax effect, especially as this phenomenon also affects instruments with coarser spatial resolution even if they do not spatially resolve the plume shadow .

A second challenge that arises with higher spatial resolution are the higher per-pixel enhancements for larger sources. As a result, the linearisation of the unit absorption spectrum around α=0 no longer holds and assumed enhancements for the calculation of the absorption spectrum have greater influences on the retrieved enhancements. Therefore, careful selection of the assumed enhancements, e.g. with the iterative approach used in this study, is essential.

Lastly, the current study shows mixed results when using the LMF introduced by . On the one hand, the proposed improvement for the detection of weak plumes was also observed in this study and lowered the detection limit even under challenging conditions. On the other hand, the LMF increased local biases in the retrieved CH₄ maps which we attribute to the amplification of noise by the log-transform in pixels with low SNR, caused by low albedo. This spatially more heterogeneous background can obscure small enhancements or produce false detections. In contrast to , we did not observe an improved performance of the LMF for large release events. The iterative MF applied in our study seems to successfully account for most non-linear absorption in pixels with large CH₄ enhancement. Therefore, further systematic analyses will be required to develop approaches that reduce or correct for this noise amplification in the LMF. Other approaches, such as WFM‑DOAS e.g., may also help better account for non‑linear effects arising from strong emission sources. However, they are computationally more expensive than the MF and tend to work better for sensors with higher spectral resolution.

As seen in Sect. , estimated emissions linearly depend on the wind speeds used. Therefore, accurate estimates of wind speeds are crucial for accurate emission estimates. Additionally, our analysis demonstrated that uncertainties in wind speeds contributed disproportionately to the uncertainty of the estimated emissions. Based on the analysis of this study, the wind speed representation error (σrepr), uncertainties in effective wind speed (σeff) and instrument precision (σinst) likely need to be revised upward. Consequently, building on the understanding of wind speed inputs (see Sect. ), future research on emission estimation from remote sensing data should prioritise methods for deriving the effective wind speed that governs plume transport (see Sect. ).