Assessment of RXD Algorithm Capability for Gas Flaring Detection through OLI-SWIR Channels

Abstract

The environment, the climate and human health are largely exposed to gas flaring (GF) effects, releasing significant dangerous gases into the atmosphere. In the last few decades, remote sensing technology has received great attention in gas flaring investigation. The Pars Special Economic Energy Zone (PSEEZ), located in the south of Iran, hosts many natural oil/gas processing plants and petrochemical industries, making this area one of the most air-polluted zones of Iran. The object of this research is to detect GF-related thermal anomalies in the PSEEZ by applying, for the first time, the Reed-Xiaoli Detector (RXD), distinguished as the benchmark algorithm for spectral anomaly detection. The RXD performances in this research field have been tested and verified using the shortwave infrared (SWIR) bands of OLI-Landsat 8 (L8), acquired in 2018 and 2019 on the study area. Preliminary results of this automatic unsupervised learning algorithm demonstrated an exciting potential of RXD for GF anomaly detection on a monthly scale (75% success rate), with peaks in the months of January and February 2018 (86%) and December 2019 (84%). The lowest detection was recorded in October 2019 (48%). Regarding the spatial distribution of GF anomalies, a qualitatively analysis demonstrated the RXD capability in mapping the areas affected by gas flaring, with some limitations (i.e., false positives) due to possible solar radiation contribution. Further analyses will be dedicated to recalibrate the algorithm to increase its reliability, also coupling L8 and Landsat 9, as well as exploring Sentinel 2 SWIR imagery, to overcome some of the observed RXD drawbacks.

1. Introduction

Gas flaring (GF) is extensively used in the petrochemical and oil/gas processing plants for removing unwanted gases in the oil and gas sector, instead of commodifying it [1,2]. Due to the environmental and human impact of gas flaring in terms of climate change, global warming and air pollution, in addition to the waste of resources and economic losses it generates [3], a global effort to track progress on gas flaring reduction will be of notability during the energy transition for clean development towards zero flaring [4,5,6,7].

Today, independent and reliable information about gas flaring can be obtained using Earth Observation (EO) satellite sensors, useful to compensate for the scarcity of ground truth data affecting most areas in the world [8]. In light of such, since the 1970s, using EO systems, offering a wide range of spectral and spatial observations, for monitoring of gas flaring activity has served as an unbiased benchmark. Numerous algorithms have been proposed, mainly aimed at detecting gas flares, quantifying their properties (such as temperature, area, and energy power) and assessing both flared volumes and flaring emissions, as described in [8].

The mean temperature of gas flares varies between 1700 and 1800 K [9,10,11]. Therefore, pixels containing a gas flare would have emission peaking in the SWIR (short-wave infrared; 1.5–3 μm) region and have different spectral behavior compared to the surrounding pixels [12]. The other electromagnetic regions also suitable for their identification span the VIS (visible, 0.4–0.7 μm), NIR (near-infrared, 0.7–1.5 μm) and MWIR (medium-wave infrared, 3–8 μm) regions [13,14,15]. The spatial extent of gas flare, whose size is of approximately 10 × 10 m² or less [16], makes them sub-pixel sources, which can be undetected in case of sensors with coarse spatial resolution (e.g., at 750 m for VIIRS, Visible Infrared Imaging Radiometer Suite) [12], while becoming the relevant hot contribution within a 30 m ground footprint, like the OLI (Operational Land Imager) one. The classical method ignores the spectral mixing phenomenon due to the very low spatial resolution of the considered data, e.g., [17].

In order to detect such small features as gas flares also in daytime conditions, the OLI sensor has been used, due to the availability of two SWIR imagery at a medium spatial resolution [18,19]. According to the consulted literature, few authors, such as [19,20,21,22,23,24,25,26], have developed techniques based on Landsat SWIR daytime imagery for identifying onshore and offshore GF related thermal anomalies, offering new interesting perspectives in this research field. Indeed, a large number of relevant papers in the literature showed the good performances of nighttime observations in recognizing gas flares [4], with a low error rate, also due to the absence of solar contamination, which affects the images acquired during daylight hours.

In this paper, the anomaly detection [27], an unsupervised learning algorithm, has been performed, for the first time, for gas flaring (GF) investigation, in contrast to standard classification tasks [21,23,25]. In recent decades, anomaly detection has considered a significant method in remote sensing interpretation. It is one of the most attractive and essential tasks for different applications and utilizations (e.g., urban planning, weather/climate forecasting, hydrological observation, military surveillance, environmental monitoring, medical analysis), where satellite data with large coverage, short temporal resolution, spectral information, and high resolution are required [28,29].

To this aim, the ReedXiaoli Detector (RXD), which refers to the problem of spotting pixels showing a curious spectral behavior when compared to all background pixels in a remotely sensed data [29]. The advantage of the RX detector is its potential in distinguishing sub-pixel targets marked by spectral behavior that appear anomalous with respect to the background without knowing the target or the background prior spectral information [30] The performance of the RXD in this field has been tested and verified over the Pars Special Economic Energy Zone (PSEEZ). It is located in southern Iran, one of the largest gas reserves in the world, where the majority of Iranian gas refineries and petrochemical plants operates. To this end, we used the multispectral L8-OLI images, binned monthly, to evaluate performance of the RXD with field data.

2. Material

2.1. Study Area

Iran is one of the top ten flaring countries, having produced 50% of global oil production and burned 17.5 billion cubic meters of natural gas in 2021, accounting for 75% of total gas flaring [31]. In 2021, Russia, Iraq and Iran were the top three countries in gas flaring in the world, and Iran accounted for 12.1% of the world’s gas flaring. Even though gas flaring has been designed for maintaining the safety of the system, in Iran, the lack of technologies and investments to capture and transport the associated gas, makes GF a top priority to be mitigated [3].

Oil/gas industries in Iran are primarily concentrated in the southern regions, particularly in the provinces of Khuzestan, Bushehr, Fars, and Kermanshah [32].

The Pars Special Economic Energy Zone (PSEEZ), located in the south of Iran in the Bushehr province on the northern coastline of the Persian Gulf, has one of the largest gas reserves in the world, where petrochemical industries, gas refineries and downstream industries have been constructed [33,34,35]. The PSEEZ zone was established in 1998 for the utilization of South Pars oil and gas resources and encouraging commercial activities in the field of oil, gas, and petrochemical industries.

In PSEEZ, there are 12 natural gas processing plants and 16 huge petrochemical complexes, containing numerous gas flare stations, that produce about 70% of the Iranian natural gas [33]. These activities have made this area one of the most polluted zones in Iran regarding the concentration of particulate matter in the air [33].

The study region is inside the PSEEZ (red box in Figure 1, left side) and consists of two zones, highlighted with boxes A and B in Figure 1 (right side), where 64 gas flares operate.

2.2. Field Observations

In the investigated area, there were 64 gas flares active between 2018 and 2019 [33]. These field points are shown in Figure 1 (green dots on the right side), where the red boxes A (center coordinates: 52°10′25″ E–27°43′48″ N) and B (center coordinates: 52°34′27″ E–27°31′23″ N) include 10 and 54 gas flares, respectively.

They are elevated, metallic vertical structures where the gas flow is fed by a derrick (from 10 to more than 100 m tall) and ignited at the top with a flame that can reach 2000 K in temperature [8]. Additionally, the gas flares used in the majority of these industries come in various types (high, medium or low pressure, liquefied petroleum gas flare, air/steam-assisted and non-assisted flares), according to data gathering from petrochemicals and oil/gas processing plants in the PSEEZ [33,34,36].

These points have been used as ground truth points to test the performance of the anomaly detection technique.

2.3. Earth Observation Data

The OLI multispectral images (from visible, B1 to short-wave infrared, B7, see

Table 1. The L8-OLI bands used in this work.

Bands	B1	B2	B3	B4	B5	B6	B7
	Coastal Aerosol	Blue	Red	Green	Near Infrared	SWIR1	SWIR2
Spectral range [μm]	0.43–0.45	0.45–0.51	0.53–0.59	0.64–0.67	0.85–0.88	1.57–1.65	2.11–2.29

), with a spatial resolution of 30 × 30 m and a temporal resolution of 16 days, covering the PSEEZ area (path 162, row 041, according to the Worldwide Reference System), were downloaded via the United States Geological Survey (https://earthexplorer.usgs.gov/ accessed on 20 April 2021).

The L8 Collection 2–Level 1 reflectance data acquired over the study area between January 2018 and December 2019 were selected and processed (

Table 2. Acquisition dates of the L8 scenes (path 162, row 041).

Year/Month	01	02	03	04	05	06	07	08	09	10	11	12	Total
2018	3,19	4,20	8,24	9,25	11,27	12,28	14,30	15,31	16	2,18	3,19	5,21	23
2019	6,22	7,23	11,27	12,28	14,30	15	1,17	2,18	3,19	5,21	6,22	8,24	23

). The digital count values were converted to Top of Atmosphere (ToA) reflectance using the radiometric rescaling coefficients stored in the metadata file. This conversion is implemented in ENVI 5.3 software.

All images (Table 2) were carefully investigated to exclude the cloudy ones from the anomaly detection analyses. For this, the meteorological data provided from the Environment Department of the PSEEZ (ground station data) for each day of the investigated dataset were taken into account and a total of 18 and 16 cloud free OLI images for 2018 and 2019, respectively, were identified. In both years, no clear images were found in April.

3. The Reed–Xiaoli Detector Method

Anomaly detection is a significant sub-division in machine learning in the era of big data and has different applications in industry, remote sensing, computer vision, and data mining [30,37,38].

The aim of anomaly detection is to distinguish targets using spectral behavior that find an anomalous pixel related to its neighboring, without knowing the target or the background prior spectral information [39]. Finally, acceptable anomaly detections do not require previous information about feature spectral behavior but a generic method for partitioning the target from background signatures. For this reason, this method has garnered a lot of attention [40,41] and is a very active field of hyperspectral and multispectral research, used, for example, to locate areas of crop stress for precision farming, to analyze oil and environmental pollution [42], and to find landmines for public safety [43].

The process of detecting anomalies includes two parts: (1) the surrounding pixels modeling and (2) using the variations betwixt the pixels of targets and the neighboring pixels to distinguish the anomalous pixels.

One of the most applicable anomaly detector is the Reed-Xiaoli (RX) [44], a second-order matched filtering algorithm that has been considered the benchmark for unsupervised anomaly detection in multispectral and hyperspectral images. The Reed-Xiaoli anomaly Detector (RXD) is capable of finding spectral differences between a target and its surrounding pixels [45], even for smaller targets. The RXD models the background as a multidimensional Gaussian distribution and estimates the deviation of a target vector from the background model [39,46,47]. When the full extent of multispectral image is used for background modeling, this is named global RX (GRX). If the RXD calculates the background through the local statistics, it is called local RX (LRX) [47]. The multivariate Gaussian distribution used by RXD assumes that the background is homogeneous (i.e., every background pixel follows a global Gaussian distribution) [39]. It calculates the overall covariance and the average value per band in a multi/hyperspectral dataset to compute the Mahalanobis distance for each pixel, which considers a metric of its anomaly. Anomalies are known as “rare events, abnormalities, deviants, or outliers” [48,49], defined by [50] as “an observation which deviates so much from the other observations as to arouse suspicions that it was calculated by a various approach”.

The GRX algorithm is expressed by Equation (1) [44,45]:

$${\sigma }_{RXD}\left(r\right)={\left(r-\mu \right)}^T{K}_{L\times L}^{-1}\left(r-\mu \right),$$

(1)

where:

-

σ_RXD(r) is the Mahalanobis distance between target to be detected and the background;

-

r is the target (i.e., the L × 1-column pixel vector in the multispectral image;

-

μ is overall background mean, given by the L—dimensional vector μ;

-

K_L×L is the test data covariance matrix, where L is the number of input spectral bands.

It is important to mention that the model assumes the data arise from two normal probability density functions with the equal covariance matrix but different averages.

Pixels showing distance values over a set threshold are assessed to be anomalous. In such a case, white pixels (like clusters) will appear on a black background over the image.

In this work, the GRX has been applied, for the first time, to detect gas flaring sources using L8-OLI observations in the above-mentioned study area. The code runs in ENVI 5.3 software, where the standard RXD is implemented (https://www.l3harrisgeospatial.com/docs/rxanomalydetection.html accessed on 11 March 2022).

4. Results and Discussion

The analyses were performed for all months in years 2018 and 2019 for pixels belonging to A and B boxes depicted in Figure 1 The results reported in the Section 4.1 and Section 4.2 refer to a sub-area shown in Figure 2a, which includes six gas flares, named as F1—F6 (Figure 2a), that are selected within the box A (Figure 1).

The data analysis consisted of three main stages:

(i)

the L8-OLI spectral values of the six GFs has been observed and discussed; the behavior of pixels with and without GFs has been plotted for the two SWIR bands (4.1 Spectral-spatial profile); the month of January 2018 has been selected as sample for their graphical representation;

(ii)

the anomalous pixels identified by RXD for each month of years 2018–2019 have been mapped over the IA (Figure 2a) (4.2 Anomaly detection);

(iii)

the operational detection rates (successful versus missed) have been computed in comparison with ground-truth available data provided by the PSEEZ for the boxes A and B (Figure 1) (4.3 Detection performance).

4.1. Spectral-Spatial Profile

Figure 3 shows the spectral variability averaged for January 2018 of six gas flares (F1–F6) in OLI bands B1–B7.

As expected, the presence of a hot pixel is reflected as an increase in the reflectance value in the SWIR spectral region at 1.6 μm (B6) and 2.2 μm (B7). Specifically, F1, F5 and F6 showed significant high value in B7 than the other flares. This is not surprising as the Google Earth picture (Figure 2b) clearly showed that these three flares are associated with larger physical structure compared to the other three flares. This may also depend on the amount of gas being flared. Indeed, flares 1, 5 and 6 are the ones with larger clusters with a higher number of white pixels (higher ToA reflectance) when compared to F2, F3 and F4 (Figure 2a). This trend was generally observed in the other months of both 2018 and 2019.

Figure 4 shows the reflectance variation in January 2018 of two SWIR bands (B6 and B7) along the green lines shown in Figure 2a.

All graphs in Figure 4 show that the signal increases when pixels that contain heat sources are crossed, with peaks belonging to F1–F6. The other ground objects are considered background, except a few (one or two) GF neighboring pixels, affected by GF activity. The reflectance values in the SWIR1 and SWIR2 B6 and B7 exhibit quite similar values, except in graphs 4 and 5, where B7 values are slightly greater than B6 values.

4.2. Anomaly Detection

The previous analysis demonstrated that the B6 and B7 L8-OLI bands are suitable for detecting anomalies related to gas flaring. These bands were stacked with each other for all months from January 2018 to December 2019. The RXD algorithm was applied to the stacked monthly values of the B6 and B7 bands for GF anomaly detection. A threshold value of 0.05 was found to provide the best detection.

Figure 5 and Figure 6 show the anomaly detection resulting maps for the six flares in 2018 and 2019, respectively, where the RXD-based anomalous pixels are marked with white clusters, while the ground-truth GF data are depicted as red points.

In 2018 (Figure 5), a thermal anomalies cluster was found in SWIR bands by the RXD algorithm for each of the six considered flares, with the only exception observed in May, when the F5 flare was not recognized.

In 2019, the RXD algorithm missed two detections, F1 in July, and F2 in August. In the other months, the RXD was able to detect all the flares.

With the exception of the above-mentioned undetected locations (i.e., F5 in May 2018, F1 in July 2019, F2 in August 2019), the other GF sites showed discernible reflectance signatures from the background. We assumed the gas flares in the three undetected cases could be temporarily out of service or burning at a much reduced combustion rate relative to their neighbors at the time of satellite passing.

The shape and size of the anomalies cluster vary with months and years. Generally, larger clusters were derived in 2019. Interestingly, the World Bank estimated that ~345 million m³ of gas were flared in 2019 as compared to ~280 million m³ burned off in the previous year (https://www.worldbank.org/en/programs/gasflaringreduction/global-flaring-data, accessed on 19 August 2020). The size of anomalies cluster is also generally larger for large flares (F1, F5, F6), probably because of the glow surrounding the flare that was treated as many individual combustion sources.

4.3. Detection Performance

The performance of the RXD in terms of monthly detection rates (successful versus missed) was assessed over the entire study area (boxes A and B in Figure 1) using the ground data of the 64 gas flares provided by the PSEEZ.

The success rate was calculated for all the months as the ratio of the correctly detected flares and the active flaring sites present in the study area:

Success Rate = (detected gas flares)/N,

(2)

where N is the number of gas flares active in the study area in 2018 and 2019 (=64).

Figure 7 shows the percentage of success rates for years 2018 (blue line in Figure 7a) and 2019 (blue line in Figure 7b), along with the number of detected and missed flares (white and blue histograms, respectively).

The success rates are greater than 70% in both 2018 and 2019, except in April of both years, when the area was cloudy, and in October 2019, when the lowest success rate was measured. The latter can be due to low intensity gas flaring or temporarily interrupted operations in that period. Further ground-truth information is required to explain these findings. Totally, the success rate (%) in 2018 is greater than in 2019.

The spatial distribution of GF anomalies from RXD is shown in Figure 8 for 2018 (box A on the left, box B on the right) and in Figure 9 for 2019 (box A on the left, box B on the right).

Figure 8 and Figure 9 showed a general consistent distribution between the RXD-detected anomalies (black pixels/clusters) and the in situ gas flares. In some cases, pixels far from the GFs systems have been flagged as anomalous from RXD: December 2019 is the most explanatory case. Focusing on this month (see the last image of Figure 9), these clusters, in part, define the perimeter of petrochemical industries and/or oil and gas processing plants (see the clusters with regular, square shape). A number of false positives affect the left side of the box B. They seem to depend on reflective solar radiation, whose effects increase in winter because of the larger solar zenith angle [19]. Indeed, these effects are more pronounced also in December 2018 (Figure 8) and February/March 2019 (Figure 9).

In other situations, the algorithm did not perform well, missing some detections. This is more evident in box B, where the majority of gas flares is localized. See, for example, the months of January–June 2018 (Figure 8), September–November 2018 (Figure 8) and the ones from May to October for 2019 (Figure 9). The flare system and the flaring conditions influence the spectral behavior of the pixel. It was observed that the majority of the undetected flares in the area are related to petrochemical flares.

According to the authors visiting the PSEEZ, the petrochemical flares stack and flame heights are significantly lower than oil/gas processing plants ones. The combustion process is completed through the injection of additional air and steam into the most of petrochemical flares. Their flames are smaller, and they may not be seen at the time of satellite acquisition. Additionally, compared to oil and gas processing plants, the rate of daily flaring in petrochemicals industries is much lower than others and most of them have a recycling system. Despite its acceptance, the RXD algorithm has drawbacks that degrade its performance in some applications [48,49,51]: (i) small size samples could not assure reliable anomaly detections; (ii) RXD often suffers from a high false positive rate; (iii) the hypothesis that the background follows a multivariate Gaussian model might not be adequate in some cases (e.g., multiple materials and textures); (iv) RXD lacks spatial recognition: each pixel is assessed independently [29]. We are planning to overcome these issues.

First, the satellite dataset will be extended in time by adding the 2020–2022 OLI observations. Second, Landsat 9 data will be included, which will halve the revisit time to 8 days, improving the chances of finding images that do not contain clouds. The potential of SWIR bands at 20 m provided by the multispectral instrument aboard the Sentinel 2 constellations, will be also explored. We will also examine additional areas affected by intense gas flaring activity. Finally, we will explore using a dynamic threshold and different bands combination to assess the RXD performance.

The comparison with current existing automatic methods for GF detection and monitoring using OLI time series, such as the one recently proposed by [26], could open interesting new scenario in this research field. Coupling detailed information on gas glaring both in space (accurate mapping of GF-affected pixels) and in time (continuous monitoring of GFs operational status) should provide a valuable contribution in quantifying the impact of flaring on air pollution and human health (at the global and country level).

5. Conclusions

We tested an unsupervised thermal anomaly detection method, the RXD algorithm, for the detection of thermal anomalies caused by flames in oil/gas processing plants and petrochemicals industries in the Pars Special Economic Energy Zone of Iran. When applied to the SWIR OLI-L8 reflectance values, in all months of 2018 and 2019, RXD showed a valuable performance (~75% success rate), providing confidence that the method can add value to current techniques in localizing and mapping areas affected by gas flaring process.

Although some limitations emerged in this study, for the presence spurious effects and missed detections, the application of the proposed detection algorithm was proved valuable for a monthly gas flaring investigation, at a middle-high spatial scale, considering the limited Landsat-8 data availability. Accuracy and sensitivity will be better analyzed by extending the RXD implementation to other satellite EO data collections as well as to regions of the world where the gas flaring assumes a paramount role for the human health. If coupled with other automatic approaches for GF investigation, this may represent an exciting and tractable tool for enabling the check of real efforts made for the elimination of gas flaring and its effects.

From a global perspective, it is necessary to monitor the top countries in the oil and gas industries according to decarbonization strategies. In this framework, satellite-based technologies and quantifying bottom-up emissions from gas flaring facilities are beneficial approaches to acquire independent and continuously updated information on the gas flaring magnitude in the space–time domain. The understanding of the importance of quantifying and monitoring of gas flaring from space by the governments, financial sponsors, operators and policy-makers will help to overcome the limitations to mitigating GF.