Reconstruction of a leaking gas cloud from a passive FTIR scanning
remote-sensing imaging system

Yunyou Hu; Yunyou Hu; Liang Xu; Xianchun Shen; Ling Jin; Hanyang Xu; Yasong Deng; Yasong Deng; Jianguo Liu; Wenqing Liu

doi:10.1364/AO.439086

1. INTRODUCTION

Chemical gas leak accidents are prone to occur in the process of production, storage, and transportation [1,2]. If toxic gas is released into the natural environment, it will cause serious damage to the health of humans. For example, during the accident of the ${{\rm{H}}_2}{\rm S}$ leakage in Kai Xian County, China, in December 2003, 243 people lost their lives [3]. Additionally, the Kab-121 platform in the Gulf of Mexico struggled from an accidental ${{\rm{H}}_2}{\rm S}$-containing natural gas leakage in 2007, which caused 21 deaths [4]. If the released gas is flammable and explosive, once it explodes, it will cause devastating damage to surrounding life and the environment. For example, the 316 hydrocarbon tank farm of a petrochemical industrial plant was involved in an accidental gas leakage and explosion in Lanzhou Province, China in January 2010, and six staff members lost their lives. Therefore, the normalized monitoring and prewarning of the source of hazardous gas leakage is of great importance.

Passive Fourier transform infrared (FTIR) remote-sensing technology [5,6] has the advantages of noncontact measurement, fast response, high sensitivity, and high resolution. Passive FTIR remote-sensing technology is combined with a two-dimensional scanning system to scan and measure toxic gases, while obtaining infrared hyperspectral data cubes [7]. The data cubes are of low spatial resolution and high spectral resolution. Low-resolution concentration-path-length (CL) images of various gas clouds are calculated by quantitative analysis methods.

The CL image of gas clouds is required to provide accurate information for calculating the gas emission flux of the chemical industrial park, the location of the leakage source with hazardous gases, the size of the toxic cloud, and the analysis of the direction of diffusion. A. Krueger and W. Stremme [8,9] used multiframe continuous column density images of gas clouds to reconstruct the propagation direction and emission rate of ${{\rm{SiF}}_4}$ and ${{\rm{SO}}_2}$ from a volcanic plume. Harig [5,6,10] analyzed chemical gases using low-resolution CL images of gas clouds. Klenk [11] used contour maps to visualize the distribution of ${{\rm{CO}}_2}$ above the chimney of a power plant. Wainner [12] scanned the leakage methane and calculated its emission flux. Xu [13] and Jiao [14,15] processed CL images of ${{\rm{SF}}_6}$ gas clouds by upsampling to analyze the gas leakage location and diffusion trend. In the reference papers above, if the low-resolution CL image of gas clouds can be converted into a high-resolution CL image of gas clouds and implemented in the analysis and calculation, the analysis of the gas cloud is more accurate. Therefore, a high-resolution CL image reconstruction method of the gas cloud leakage applied to the passive FTIR scanning remote-sensing imaging system is proposed. There are two major difficulties in reconstruction: (a) the CL image of a gas cloud has low resolution, and (b) the CL image of a gas cloud is affected by wind direction and wind speed, making the leaking gas cloud more difficult to trace.

The imaging process of the CL image of a gas cloud is complicated. It is affected by the distance of the target gas cloud and the passive FTIR scanning remote-sensing system’s own factors (the infrared detector is a unit detector, and the infrared field of view is much larger than the visible light field of view). Therefore, the column density image of the gas cloud is a low-resolution image. To address the problem of low-resolution images, many image interpolation algorithms have been proposed. These algorithms can be roughly divided into two categories in terms of technical ideas: interpolation methods based on local data (local-data-based) [16–20] and example-based interpolation methods [21–24]. The low-resolution CL images of gas clouds collected by the FTIR remote-sensing imaging system do not have standard high-resolution images as a reference. The interpolation method based on examples cannot be applied to this system. Compared with the advanced interpolation algorithm based on local data [17], bicubic convolution interpolation [25] has the lowest time complexity, and its interpolation accuracy is roughly equivalent [17–20,24,26]. Bicubic interpolation can achieve better results when applied to passive FTIR remote-sensing imaging systems.

When leaking, the gas will diffuse outward in the form of a Gaussian diffusion model [27–29] to form a leaking gas cloud. Affected by the surrounding wind direction and wind speed, the gas cloud will spread more rapidly. Therefore, the measured gas cloud image consists of a leaking gas cloud and several diffused gas clouds. The leaking gas cloud has features such as a maximum gas CL, a concentrated trend, and a large size. It can be accurately detected with these features. To obtain the leaking gas cloud and assess the source of the leak, a method to reconstruct CL images of leakage gas clouds is proposed, which includes four steps: upsampling, image segmentation, morphological processing, and leak evaluation. The correctness of the method is tested by simulation, and then the remote measurement experiment is carried out with ${{\rm{SF}}_6}$ as the tracer gas. The results indicate that the proposed method effectively realizes the reconstruction of the high-resolution CL image of the gas cloud leakage, and the reconstruction method has excellent accuracy and robustness in real time. The effect of the reconstructed leaked gas cloud is crucial, the plume trajectory is clear, and the source of the leakage is obvious.

2. METHODOLOGY

A. Imaging Method for the Low-Resolution CL Image of a Gas Cloud

Passive remote sensing of hazardous gases is based on the analysis of the infrared absorption and radiation characteristics of various molecules in the gas cloud. Figure 1 is a simplified three-layer radiation transmission model. The infrared radiation received by the unit detector contains the spectral characteristics of the background, the threat gas cloud, and the atmosphere. Radiation transmission theory [30] is used to describe the propagation process of radiation in the atmosphere, where the atmosphere and the threatening gas cloud layer can be regarded as homogeneous layers. Radiation from the background layer, such as the sky, the ground, or a building, passes through the target gas cloud and atmosphere to reach the infrared detector.

Fig. 1. Three-layer radiation transmission model for passive FTIR remote sensing of polluted gas clouds.

Download Full Size | PDF

In this model, the radiance ${L_1}(\nu)$ measured by the spectrometer is

(1)$$\begin{split}{L_1}(\nu) &= (1 - {\tau _1}(\nu)){B_1}(\nu)\\&\quad + {\tau _1}(\nu)[(1 - {\tau _2}(\nu)){B_2}(\nu) + {\tau _2}(\nu){L_3}(\nu)],\end{split}$$

where $v$ is the wavenumber of the FTIR spectra, ${\tau _{\rm{i}}}(\nu)$ is the transmittance of layer i, ${B_{\rm{i}}}(\nu)$ is the radiance of a blackbody at the temperature of layer i, and ${L_3}(\nu)$ is background radiation. The contribution of scattering is negligible. If the transmittance of the first layer is equal to 1, Eq. (1) can be simplified, and the transmittance of the gas cloud can be expressed by

(2)$${\tau _2}(\nu) = \frac{{{L_1}(\nu) - {B_2}(\nu)}}{{{L_3}(\nu) - {B_2}(\nu)}}.$$

If the radiance of the background and the temperature of the gas cloud are known, it is possible to calculate the transmittance. In many cases, it is not possible to measure the background spectrum. To meet the application requirements of online measurement, this paper uses a real-time background extraction method from the measured spectrum [6,15]. The blackbody radiance spectrum of the gas cloud can be estimated by the brightness temperature in the opaque spectrum of the atmosphere, and the background spectrum can be estimated by the brightness temperature of the transparent spectrum of the atmosphere [15]. After acquiring the transmittance spectrum of the gas cloud, the Beer–Lambert law [Eq. (3)] can be used to invert the CL of the gas target [10],

(3)$$\tau (\nu) = \exp (- \alpha (\nu) \cdot {\rm{CL}}),$$

where CL is the CL product and $\alpha (\nu)$ is the absorption coefficient of gas molecules.

Combining FTIR technology and scanning, we can obtain an infrared hyperspectral data cube with a low spatial resolution. We assume that the dimensions of the data cube are (m, n, s), where (m, n) is the spatial dimension and s is the spectral dimension. The inversion of the gas CL is performed by an iterative nonlinear least-squares algorithm [31], which can be expressed as

(4)$${{\rm{CL}}_{i,j}} = \arg \min {\sum\limits_\nu {({{\tau ^\prime}_{i,j}}(\nu) - {\tau _{i,j}}(\nu))} ^2},$$

where ${\rm{i}} = 1,2,\ldots m;j = 1,2,\ldots n$, ${{\rm{CL}}_{i,j}}$ is the CL at coordinates (i, j). ${\tau ^\prime _{i,j}}(\nu)$ is the transmittance spectrum obtained after multiple iterations of the spectrum in the HITRAN database. ${\tau _{i,j}}(\nu)$ is the transmittance spectrum calculated by Eq. (2). We arrange the ${{\rm{CL}}_{i,j}}$ of the gas cloud obtained by quantitative calculation as a matrix in the scanning order. Therefore, the low-resolution CL image ${{\rm{X}}_L}$ of the gas cloud is defined by

(5)$${{\rm{X}}_{\rm{L}}} = \left[{\begin{array}{*{20}{c}}{{{\rm{CL}}_{1,1}}}&{{{\rm{CL}}_{1,2}}}&{\ldots}&{{{\rm{CL}}_{1,n}}}\\{{{\rm{CL}}_{2,1}}}&{{{\rm{CL}}_{2,2}}}&{\ldots}&{{{\rm{CL}}_{2,n}}}\\{\ldots}&{\ldots}&{\ldots}&{\ldots}\\{{{\rm{CL}}_{m,1}}}&{{{\rm{CL}}_{m,2}}}&{\ldots}&{{{\rm{CL}}_{m,n}}}\end{array}} \right],$$

where (m, n) is the dimensional size of the CL image of the gas cloud.

B. Reconstruction of a High-Resolution CL Image of a Leaking Gas Cloud

When leaking, the gas will form a leaking gas cloud in the form of Gaussian diffusion. The leaking gas cloud has the characteristics of a high CL, a concentrated trend, and a large size. If affected by the surrounding wind direction and wind speed, the leaking gas cloud will spread rapidly, forming a gas cloud image with the leaking gas cloud as the main part and the diffusion gas cloud as the supplement. The diffusion gas cloud has a low CL value and is of small size. The reconstruction method of the leaking gas cloud is based on the above assumptions. Figure 2 shows the reconstruction method of the gas cloud leakage, which is described in detail below.

Fig. 2. Reconstruction method of the CL image of the leakage gas cloud.

Download Full Size | PDF

Fig. 3. Passive FTIR scanning remote-sensing imaging system. (a) Experimental equipment; (b) schematic diagram of two-dimensional scanning; (c) human-computer interaction interface.

Download Full Size | PDF

The interpolation algorithm is used to improve the resolution of the CL image of the gas cloud. The essence of bicubic convolution interpolation is to use the sum of the weighted convolutions of the 16 neighboring pixels of the image as a new pixel value. The convolution kernel is the bicubic kernel function ${\rm{K}}({\rm{t}})$, which often takes ${-}{0.5}$, and the function ${\rm{K}}({\rm{t}})$ is defined by

(6)$$K(t) = \left\{{\begin{array}{*{20}{c}}{(a + 2){{\left| t \right|}^3} - (a + 3){{\left| t \right|}^2} + 1}\\{a{{\left| t \right|}^3} - 5a{{\left| t \right|}^2} + 8a\left| t \right| - 4a}\\0\end{array}} \right.\begin{array}{*{20}{c}}{{\rm when}\;\left| t \right| \le 1}\\{{\rm when}\; 1 \lt \left| t \right| \le 2}\\{\rm else}\end{array}.$$

The low-resolution CL image ${X_L}$ of the gas cloud is magnified by M times and becomes ${X_H}$ after interpolation. The result of interpolation ${X_H}$ at point $(x,y)$ is as follows:

(7)$$\begin{split}{{\rm{X}}_H}(x,y) & = {X_L}([i] + u,[j] + v)\\& = \sum\limits_{r = - 1}^2 \sum\limits_{c = - 1}^2 {X_L}([i] + r,[j] + c)\\&\quad \cdot K(r - u) \cdot K(c - v), \end{split}$$

where $i = \frac{x}{M},j = \frac{y}{M},u = i - [i],v = j - [j]$, and $[*]$ is the rounding function. The image pixel value is affected by the accuracy of the column density inversion algorithm. Therefore, the maximum of the gas cloud noise-equivalent concentration-path-length (NECL) image [5,14] is used as the segmentation threshold. NECL is expressed by

(8)$${\rm{NECL}} = - \frac{{\lg \left({1 - \frac{{\rm NESR}}{{\left| {{B_{{2}}}({\nu _{{\rm{peak}}}}) - {L_3}({\nu _{\rm{peak}}})} \right|}}} \right)}}{{\alpha ({\nu _{{\rm{peak}}}})}},$$

where NESR is the noise-equivalent spectral radiance, ${\nu _{{\rm{peak}}}}$ is the wavenumber at the absorption peak, ${L_3}$ is the background spectrum, and ${B_{{2}}}$ is the gas cloud spectrum. $\alpha ({\nu _{{\rm{peak}}}})$ is the absorption coefficient of the gas cloud. The maximum NECL is used as the segmentation threshold of the CL image of the gas cloud to generate the mask, and the mask is defined by

(9)$${{\rm{M}}_{{0}}}({X_H}) = \left\{{\begin{array}{*{20}{c}}1&{{X_H}(i,j) \ge {\rm{NECL}}}\\0&{else}\end{array}} \right..$$

There are small connections or burrs in the mask image at the gas cloud boundary. Through image morphology processing [Eq. (10)], we first erode, then dilate [32,33]. The purpose of this step is to eliminate small gas clouds, separate the gas clouds at the small connection place, and smooth the boundary of the larger gas clouds. This step is as follows:

(10)$${{\rm{M}}_{{1}}}= ({{\rm{M}}_{{0}}} \ominus {\rm{B}}) \oplus {\rm{B,}}$$

where ${\rm{B}} = \left[{\begin{array}{*{20}{c}}0&0&1&0&0\\0&1&1&1&0\\1&1&1&1&1\\0&1&1&1&0\\0&0&1&0&0\end{array}}\right]$, $\ominus$ erode operation, $\oplus$ dilate operation, and ${{\rm{M}}_1}$ is a new template containing several gas clouds. After the morphological processing of the image, a small number of edge pixels becomes lost in the image; however, a better segmented template, ${{\rm{M}}_1}$, is obtained. The connected components labeling method [32] is applied to the ${{\rm{M}}_1}$ template, and the pixel in the same gas cloud is labeled with the same label number to facilitate the statistics of image characteristics. The mask of each gas cloud is obtained by the label, and n is the number of labels. The pixel value of the gas cloud image contains only one type of gas, and the gas diffuses rapidly in Gaussian form after the leakage. When the wind speed is low, a leaking gas cloud with the largest size, a high CL, and concentrated trend is formed. It may also be affected by the harsh environment, and these three characteristics exist within three gas clouds. Therefore, a number of gas clouds after segmentation are prioritized in the order of gas cloud size, maximum CL of a gas cloud, and the average CL of a gas cloud. The gas cloud is obtained as follows:

(11)$${{{X}}_{\rm gas\_cloud}}=\left\{ {{{X}}_{H}}\cdot {{M}_{s}},{{{X}}_{H}}\cdot {{M}_{mcl}},{{{X}}_{H}}\cdot {{M}_{acl}} \right\},$$

where ${M_s}$ is the gas cloud template with priority on gas cloud size; ${M_{{mcl}}}$ is the gas cloud template with priority on the maximum CL of gas cloud; ${M_{{acl}}}$ is the gas cloud template with priority on the average CL of the gas cloud; and $\{* \}$ is a set.

C. Assessment of a High-Resolution CL Image of a Reconstructed Leaking Gas Cloud

Intersection over union (IoU) [34]: The ratio of the intersection of the segmented gas cloud area (${{\rm{C}}_1}$) and the real gas cloud area (${{\rm{C}}_2}$) to the union portion can be calculated here,

(12)$${\rm{IoU}} = \frac{{N({C_1} \cap {C_2})}}{{N({C_1} \cup {C_2})}},$$

where $N(*)$ is for counting the number of pixels in the area. The closer the IoU is to 1, the better the segmentation effect.

3. SIMULATION AND EXPERIMENT

A. Passive FTIR Scanning Remote-Sensing Imaging System

The passive FTIR scanning remote-sensing imaging system is shown in Fig. 3(a), and Table 1 shows the system parameters. The system consists of an FTIR spectrometer, scanning mirror, reflective telescope, video camera, a 2-DOF pan-tilt, and a data acquisition and processing system. Figure 3(b) is a schematic diagram of a two-dimensional scan while the system is working. In the scanning process, infrared information and background images are collected simultaneously. The field of view in the infrared telescope is 7.5 mrad, and the single-pixel field of view of the video camera is approximately 0.553 mrad. Figure 3(c) is the system software interactive interface. The background image and the false color with column density are superimposed and displayed in the center area. For the visualization of pollutant clouds, the scanning mirror is set in order within the field of regard. The size of the monitoring field, the spatial resolution and the number of spectra per pixel can be set by the user. When the system is working, the scanning system cooperates with an FTIR spectrometer to scan and measure the gas cloud according to the sampling array preset by the user. The infrared radiance spectrum of each pixel is obtained by the data acquisition and processing system and is then transmitted to the PC for identification, quantification, and display. At the same time, the scanning mirror is pointed to the next pixel. Finally, the false color with the gas CL is combined with the visible light background image to synthesize a gas cloud image.

Table 1. Parameters of the Passive FTIR Scanning Remote-Sensing Imaging System

View Table | View all tables in this article

B. Simulation

To evaluate the effectiveness, robustness, and computational efficiency of the algorithm, the CL image of the gas cloud used in the simulation is acquired by superimposing multiple two-dimensional Gaussian distributions [27]. The two-dimensional Gaussian distribution function is as follows:

(13)$${\rm{CL}}(x,y|A,{x_0},{y_0},{\sigma _x},{\sigma _y}) = A\exp \left[{- \frac{{{{(x - {x_0})}^2}}}{{2{\sigma _x}^2}} - \frac{{{{(y - {y_0})}^2}}}{{2{\sigma _y}^2}}} \right],$$

where ${\rm{CL}}(x,y|A,{x_0},{y_0},{\sigma _x},{\sigma _y})$ is the CL of the gas cloud, A is the peak CL, $({x_0},{y_0})$ is the peak coordinate, and ${\sigma _x}$ and ${\sigma _y}$ are the standard deviations of the peaks in the $x$ and $y$ directions, respectively.

We randomly generate 500 CL images of the gas cloud. The parameters are shown in Table 2, and each image is composed of five gas clouds added together. The peak, position, and size of each gas cloud are randomly generated within a range. Enter $x$ from ${-}{{4}}$ to 4, and enter $y$ from ${-}{{4}}$ to 4, with an interval of 0.1, and the resulting image size is (80,80). Consistent with the way the remote-sensing system collects infrared radiation, the downsampling method is average downsampling, and the sampling factor is 0.25. Five hundred low-resolution images of the gas cloud were generated by the simulation.

Table 2. Parameter Range of Simulated Gas Cloud Image

View Table | View all tables in this article

C. Field Experiment

Figure 4 shows the field experiment scene in Hefei on February 2, 2021 taken by the video camera of the passive FTIR scanning remote-sensing imaging system. A bottle of ${{\rm{SF}}_6}$ gas is placed in the red frame, the direction of the red arrow is the direction of the gas outlet, and the blue box is the scan area that is set manually. The detection background is an environmental background at a distance of 200 m, which includes buildings and vegetation. The target gas is ${{\rm{SF}}_6}$ released from the pressure vessel. Before the experiment, the passive FTIR scanning remote-sensing imaging system is calibrated with blackbody radiation in the laboratory, and the calibration temperature is 30°C and 60°C. After deducting the absorption of water vapor and ${{\rm{CO}}_2}$, the calibration coefficient is calculated by a two-point calibration method.

Fig. 4. Scene of the field experiment.

Download Full Size | PDF

The experiment is divided into three phases: measurement before deflation, measurement during deflation, and measurement after deflation. The specific detection details of a single-column density image are as follows: The pixel array of the area to be measured is set to ${{9}} \times {{11}}$, and the interferogram is collected in line order. After collecting the interferogram of a pixel area in the target space, the system calculates the corresponding ${{\rm{SF}}_6}$ transmittance spectrum and inverts the ${{\rm{SF}}_6}$ CL by the nonlinear least-square method. At the same time, the scanning system drives the scanning mirror to point to the next pixel and performs detection, forming a complete column density low-resolution image.

4. RESULTS AND DISCUSSION

A. Simulation Results and Discussion

Both simulation and experimental calculations are implemented in MATLAB. The proposed method performs a high-resolution reconstruction of the leaking gas cloud on 500 low-resolution images generated by simulation. Determined by the ratio of the camera’s field of view to the scanning mirror’s field of view, the upsampling factor used in passive FTIR scanning remote-sensing imaging systems should not be too large, and a more appropriate value is 4 times. We select $0.5\;{{\rm ppm}} \cdot {\rm{m}}$ as the segmentation threshold of the simulated image, which is close to the maximum NECL of the experiment in Section 4.B. A total of 500 low-resolution gas cloud images are used as the input of the proposed algorithm. The average calculation time of each image is 9.85 ms, which shows a strong real-time performance. Figure 5 is the simulation result. The IoU varies within the range of 0.97 to 1, and the leaking gas cloud can be segmented accurately. The number of pixels in a single simulated image is 6400, and the pixels of the leaking gas cloud vary from 1000 to 3000. The IoU value in Fig. 5 shows that the three features proposed in this paper accurately find all leaking gas clouds from the 500 simulation images.

Fig. 5. Size of the leaking gas cloud and the curve of IoU.

Download Full Size | PDF

Fig. 6. Measured spectrum, extracted background spectrum, and extracted gas layer spectrum.

Download Full Size | PDF

Fig. 7. Transmittance spectrum obtained by the method in Section 2.

Download Full Size | PDF

Table 3. CL Distribution of the ${{\rm{SF}}_6}$ Gas Cloud at 15:54:07

View Table | View all tables in this article

In this simulation, five gas clouds are added to a gas cloud image according to the spatial distribution, which simulates the influence of the environment, such as wind speed and direction. The simulated image resolution is (20,20), which also matches the measurement image size of the instrument well. The gas cloud fitted by the Gaussian distribution is consistent with the diffusion pattern of the measured gas cloud. The results obtained in this simulation mode are close to the actual situation, and it has excellent practicability and robustness.

B. Experiment Results and Discussion

1. Imaging Process of Low-Resolution CL Images of Gas Clouds

The radiation is measured by a passive FTIR scanning remote-sensing imaging system, and eight interferograms are added to one scan. Figure 6 is a measured spectrum of the experiment. The background spectrum and the gas layer spectrum are extracted from the measured spectrum by the method introduced in Section 2.A. The background spectrum is the blackbody radiation at 286.84 K, and the gas layer spectrum is the blackbody radiation at 237.65 K. The transmittance spectrum shown in Fig. 7 is calculated by Eq. (2).

With reference to the standard gas database and Beer–Lambert law, CL inversion is performed on the transmittance spectrum using the nonlinear least-square method. The inversion band of ${{\rm{SF}}_6}$ is ${{912 \!-\! 968}}\;{\rm{c}}{{\rm{m}}^{- 1}}$, and the results are arranged in a matrix (Table 3) according to Eq. (5).

The blackbody radiation ${{\rm{L}}_{45}}$ is measured at 45°C, and the blackbody radiation ${{\rm{B}}_{45}}$ is calculated at 45°C according to the Planck function. The measured spectrum ${{\rm{L}}_{45}}$ is subtracted from the calculated spectrum ${{\rm{B}}_{45}}$ to obtain the differential spectrum ${{\rm{LB}}_{45}}$. ${\rm{NESR}} = 7.42 \times {10^{- 9}}{\rm{W/(c}}{{\rm{m}}^{{2}}} \cdot {\rm{sr}} \cdot {\rm{c}}{{\rm{m}}^{- 1}}{\rm{)}}$ is calculated by the inversion band in the differential spectrum ${{\rm{LB}}_{45}}$. The maximum absorption coefficient of ${{\rm{SF}}_{6}}$ in the inversion band is $0.0323\;{\rm{pp}}{{\rm{m}}^{- 1}} \cdot {{\rm{m}}^{- 1}}$ (obtained from the commercial standard spectrum). We insert the spectrum of each pixel into Eq. (8) to calculate and obtain NECL from 0.036 to $0.451\;{\rm{ppm}} \cdot {\rm{m}}$.

Table 3 shows the CL distribution of the gas cloud when ${{\rm{SF}}_6}$ is first released. According to the registration method of the infrared field of view and the camera field of view, the values in Table 3 are superimposed on the background image. Figure 8(a) shows the false-color image of the gas cloud and the background image. It is easy to find that the false-color image in Fig. 8(a) is a low-resolution image, and the high values are roughly distributed at the gas cylinder outlet. The leaking gas cloud is split into several gas clouds under the influence of wind speed and direction. Although the leaking gas cloud can be detected with the eyes, there are also many interfering gas clouds. This is inconvenient for quickly analyzing the location of the leakage.

Fig. 8. Reconstruction process of the leaking gas cloud image. (a) Low-resolution CL image; (b) high-resolution CL image with bicubic interpolation; (c) mask segmented by threshold; (d) mask processed by morphology; (e) gas clouds divided by a mask; (f) leaking gas cloud selected by the evaluation criteria.

Download Full Size | PDF

2. High-Resolution Reconstruction of the ${{\rm{SF}}_6}$ Leaking Gas Cloud

We let the ${{\rm{SF}}_6}$ CL distribution in Table 3 be the input of the proposed algorithm. The ${{\rm{SF}}_6}$ low-resolution column density image [Fig. 8(a)] is subjected to fourfold bicubic convolution interpolation [Fig. 8(b)]. Let the maximum NECL ($0.45\,{\rm{ppm}} \cdot {\rm{m}}$) be used as the threshold of image segmentation to generate a mask [Fig. 8(c)]. Through image morphology processing (eroding first and then dilating), the small gas cloud is removed, the gas cloud is separated at its weak connection part, and the boundary of the larger gas cloud is smoothed. Figure 8(d) shows that this step effectively separates the gas cloud. As shown in Table 4, taking the maximum NECL as the segmentation threshold, five gas clouds are obtained. Among them, gas cloud 4 has the largest size, the largest CL value, and the largest average CL. Gas cloud 4 satisfies the three characteristics at the same time, and thus, it can be determined that the leaking gas cloud is gas cloud 4. The entire algorithm process takes 4.05 ms, which meets the real-time requirements of the passive FTIR remote-sensing imaging system. The IoU is 0.967, indicating that the segmentation accuracy of the leaking gas cloud is high. After comparing Fig. 4 with Fig. 8(f), it is evident that gas cloud 4 is distributed on the plume path at the mouth of the gas cylinder. Therefore, gas cloud 4 can be considered a leaking gas cloud.

Table 4. Gas Cloud Parameters Corresponding to Fig. 8(e)

View Table | View all tables in this article

Figure 9 shows the relationship between the segmentation threshold (NECL) and the size of the leaking gas cloud, as well as the relationship between the segmentation threshold and the IoU. This shows the feasibility of the maximum NECL as the segmentation threshold of the gas cloud images. As the NECL value increases, the size of the leaking gas cloud gradually decreases in the interval below the maximum NECL, which indicates that there are multiple gas clouds stuck together below the maximum NECL ($0.45\;{\rm{ppm}} \cdot {\rm{m}}$). After the NECL reaches $0.45\;{\rm{ppm}} \cdot {\rm{m}}$, the changes in gas cloud size tend to be stable, which shows that the leaking gas cloud is separated from other diffused gas clouds. When the NECL reaches $0.9\;{\rm{ppm}} \cdot {\rm{m}}$, there is a step change in the size of the leaking gas cloud, which separates the low-value area at the lower right of gas cloud 4. Regardless of how the threshold (NECL) changes, the IoU fluctuates in a stable range, which shows that the proposed method is robust and accurate.

Fig. 9. Size change and accuracy assessment diagram of the leaking gas cloud under different segmentation thresholds.

Download Full Size | PDF

Whether it is a leaking gas cloud or a diffused gas cloud, the gas is diluted to the surroundings from a high concentration location. Therefore, we need to simply use the threshold to separate each gas cloud. Choosing a larger threshold will reduce the size of the leaking gas cloud. Choosing a smaller threshold will make the leaking gas clouds stick together. The maximum NECL is used as our segmentation threshold; the CL value below the maximum NECL can be ignored. Additionally, adhesion between gas clouds can occur when the maximum NECL is a threshold. The remaining adhesion can be resolved by morphological processing, and several relatively complete gas clouds can be obtained. The leaking gas cloud is detected from multiple gas clouds by Eq. (11), but the evaluation criterion is valid when there is only one leak source. In addition, even when there are multiple sources of leakage, only the strongest source can be found.

5. CONCLUSION

The CL image of the gas cloud measured by the passive FTIR remote-sensing imaging system has a low resolution, and the diffusion trend of the leaking gas cloud is affected by the wind speed and direction, which makes it difficult to obtain a more accurate analysis. A reconstruction method more suitable for finding the leaking gas cloud is proposed. The effectiveness, real-time reliability, and robustness of the leaking gas cloud reconstruction method are demonstrated by simulation and field experiments. The leaking gas cloud reconstruction method has been successfully applied to passive FTIR scanning remote-sensing imaging systems, providing high-quality and high-resolution visualization images for analyzing gas diffusion, leak location, gas cloud size, and chemical emission flux calculations. However, the method proposed in this paper can only be applied when there is one leakage point. The next step is to study the leaking gas cloud reconstruction method when there are multiple leakage points.

Funding

National Key Research and Development Program of China (2019YFF0303400); Key Research Program of Frontier Science, Chinese Academy of Sciences (QYZDY-SSW-DQC016); National Natural Science Foundation of China (41941011).

Acknowledgment

The authors would like to thank anonymous reviewers for their insightful comments and constructive suggestions.

Disclosures

The authors declare no conflicts of interest.

Data Availability

Data underlying the results presented in this paper are not publicly available at this time but may be obtained from the authors upon reasonable request.

REFERENCES

1. Y. Cao, A. Liu, M. Chen, Y. Jiang, S. Wang, and Y. Bai, “Safety analysis of an offshore platform for leakage and deflagration accidents from adjacent oil and gas storage and transportation units,” Ships Offshore Struct. 16, 815–826 (2021). [CrossRef]

2. M. J. Jafari, I. Mohammadfam, and E. Zarei, “Analysis and simulation of severe accidents in a steam methane reforming plant,” Int. J. Occup. Hyg. 6, 120–130 (2014).

3. D. Yang, G. Chen, J. Shi, Y. Zhu, and Z. Dai, “A novel approach for hazardous area identification of toxic gas leakage accidents on offshore facilities,” Ocean Eng. 217, 107926 (2020). [CrossRef]

4. D. Yang, G. Chen, J. Shi, and X. Li, “Effect of gas composition on dispersion characteristics of blowout gas on offshore platform,” Int. J. Naval Arch. Ocean Eng. 11, 914–922 (2019). [CrossRef]

5. R. Harig and G. Matz, “Toxic cloud imaging by infrared spectrometry: a scanning FTIR system for identification and visualization,” Field Anal. Chem. Technol. 5, 75–90 (2001). [CrossRef]

6. M. Grutter, R. Basaldud, E. Flores, and R. Harig, “Optical remote sensing for characterizing the spatial distribution of stack emissions,” in Advanced Environmental Monitoring (Springer, 2008), pp. 107–118.

7. D. G. Manolakis, S. E. Golowich, and R. S. DiPietro, “Long-wave infrared hyperspectral remote sensing of chemical clouds: a focus on signal processing approaches,” IEEE Signal Process. Mag. 31(4), 120–141 (2014). [CrossRef]

8. A. Krueger, W. Stremme, R. Harig, and M. Grutter, “Volcanic SO₂ and SiF₄ visualization using 2-D thermal emission spectroscopy. Part 2. Wind propagation and emission rates,” Atmos. Meas. Tech. 6, 47–61 (2013). [CrossRef]

9. W. Stremme, A. Krueger, R. Harig, and M. Grutter, “Volcanic SO₂ and SiF₄ visualization using 2-D thermal emission spectroscopy. Part 1. Slant-columns and their ratios,” Atmos. Meas. Tech. 5, 275–288 (2012). [CrossRef]

10. R. Harig, “Passive remote sensing of pollutant clouds by Fourier-transform infrared spectrometry: signal-to-noise ratio as a function of spectral resolution,” Appl. Opt. 43, 4603–4610 (2004). [CrossRef]

11. U. Klenk, E. Schmidt, and A. Beil, “Scanning infrared remote sensing system for detection, identification and visualization of airborne pollutants,” in Atmospheric and Biological Environmental Monitoring (Springer, 2009), pp. 51–62.

12. R. T. Wainner, N. F. Aubut, M. C. Laderer, and M. B. Frish, “Scanning, standoff TDLAS leak imaging and quantification,” Proc. SPIE 10210, 1021006 (2017). [CrossRef]

13. L. Xu, M. G. Gao, J. G. Liu, Y. Jiao, M. C. Feng, J. J. Tong, and S. Li, “Remote sensing of chemical gas cloud emission by passive infrared scanning imaging system,” Proc. SPIE 8534, 85340M (2012). [CrossRef]

14. Y. Jiao, L. Xu, M. G. Gao, L. Jin, J. J. Tong, S. Li, and X. L. Wei, “Investigation of the limit of detection of an infrared passive remote sensing and scanning imaging system for pollution gas,” Spectrosc. Spectral Anal. 33, 2617–2620 (2013).

15. Y. Jiao, L. Xu, M. G. Gao, L. Jin, J. J. Tong, S. Li, and X. L. Wei, “Real-time data processing of remote measurement of air pollution by infrared passive scanning imaging system,” Acta Phys. Sinica 62, 140705 (2013). [CrossRef]

16. X. Li and M. T. Orchard, “New edge-directed interpolation,” IEEE Trans. Image Process. 10, 1521–1527 (2001). [CrossRef]

17. Z. F. Lu and B. J. Zhong, “Image interpolation with predicted gradients,” Acta Automat. Sinica 44, 1072–1085 (2018).

18. Z. Wei and K. K. Ma, “Contrast-guided image interpolation,” IEEE Trans. Image Process. 22, 4271–4285 (2013). [CrossRef]

19. W. Ye and K. K. Ma, “Convolutional edge diffusion for fast contrast-guided image interpolation,” IEEE Signal Process. Lett. 23, 1260–1264 (2016). [CrossRef]

20. L. Zhang and X. L. Wu, “An edge-guided image interpolation algorithm via directional filtering and data fusion,” IEEE Trans. Image Process. 15, 2226–2238 (2006). [CrossRef]

21. C. Huang, X. Q. Ding, and C. Fang, “Single-image super-resolution via adaptive joint Kernel regression,” in British Machine Vision Conference (2013).

22. J. J. Huang, W. C. Siu, and T. R. Liu, “Fast image interpolation via random forests,” IEEE Trans. Image Process. 24, 3232–3245 (2015). [CrossRef]

23. J. J. Jiang, X. Ma, C. Chen, T. Lu, Z. Y. Wang, and J. Y. Ma, “Single image super-resolution via locally regularized anchored neighborhood regression and nonlocal means,” IEEE Trans. Multimedia 19, 15–26 (2017). [CrossRef]

24. Y. Romano, M. Protter, and M. Elad, “Single image interpolation via adaptive nonlocal sparsity-based modeling,” IEEE Trans. Image Process. 23, 3085–3098 (2014). [CrossRef]

25. D. Q. Liu, G. Q. Zhou, X. Zhou, C. Y. Li, and F. Wang, “FPGA-based on-board cubic convolution interpolation for spaceborne georeferencing,” Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci. XLII-3/W10, 349–356 (2020). [CrossRef]

26. S. Mallat and G. Yu, “Super-resolution with sparse mixing estimators,” IEEE Trans. Image Process. 19, 2889–2900 (2010). [CrossRef]

27. C. L. Deng, J. J. Tong, M. G. Gao, X. X. Li, Y. Li, X. Han, and W. Q. Liu, “Simulation of tomographic reconstruction algorithms for open-path Fourier transform infrared spectroscopy,” Acta Opt. Sinica 39, 0707001 (2019). [CrossRef]

28. Y. B. Ren, Y. Li, J. D. Wang, X. F. Wang, B. P. Liu, L. M. Zhang, and L. Zhang, “Reconstruction of air contaminant concentration distribution in a two-dimensional plane by computed tomography and remote sensing FTIR spectroscopy,” J. Environ. Sci. Health A 40, 571–580 (2005). [CrossRef]

29. L. A. Todd, M. Ramanathan, K. Mottus, R. Katz, A. Dodson, and G. Mihlan, “Measuring chemical emissions using open-path Fourier transform infrared (OP-FTIR) spectroscopy and computer-assisted tomography,” Atmos. Environ. 35, 1937–1947 (2001). [CrossRef]

30. R. M. Goody and Y. L. Yung, Atmospheric Radiation: Theoretical Basis (Oxford University, 1995).

31. T. E. L. Smith, M. J. Wooster, M. Tattaris, and D. W. T. Griffith, “Absolute accuracy and sensitivity analysis of OP-FTIR retrievals of CO₂, CH₄ and CO over concentrations representative of ‘clean air’ and ‘polluted plumes,’,” Atmos. Meas. Tech. 4, 97–116 (2011). [CrossRef]

32. O. Kalentev, A. Rai, S. Kemnitz, and R. Schneider, “Connected component labeling on a 2D grid using CUDA,” J. Parallel Distrib. Comput. 71, 615–620 (2011). [CrossRef]

33. P. Li, T. Shi, A. Lu, and B. Wang, “Quantum circuit design for several morphological image processing methods,” Quantum Inf. Process. 18, 364 (2019). [CrossRef]

34. Z. Zheng, P. Wang, W. Liu, J. Li, R. Ye, and D. Ren, “Distance-IoU loss: faster and better learning for bounding box regression,” in AAAI Conference on Artificial Intelligence (2020), pp. 12993–13000.

Parameter	Value
Spectral range	$600 - 1800 c m^{- 1}$
Resolution for remote sensing	$4 c m^{- 1}$
Field of view with telescope	7.5 mrad
Field of regard	$360^{\circ} \times 60^{\circ}$
Scan rate ( $Δ \bar{ν} = 4 c m^{- 1}$ )	10 spectra/s
Angular resolution of the scanner	0.01°

	Parameter
Number	A	$x_{0}$	$y_{0}$	$σ_{x}$	$σ_{y}$
1	(4,5)	(−4,4)	(−4,4)	(1.5,3.0)	(1.5,3.0)
2	(3,4)	(−4,4)	(−4,4)	(1.0,2.5)	(1.0,2.5)
3	(2,3)	(−4,4)	(−4,4)	(0.5,2.0)	(0.5,2.0)
4	(1,2)	(−4,4)	(−4,4)	(1.0,1.5)	(1.0,1.5)
5	(0,1)	(−4,4)	(−4,4)	(0.0,1.0)	(0.0,1.0)

	Column
Row	1	2	3	4	5	6	7	8	9	10	11
1	0.25	0.06	0.32	0.64	0.92	0.34	0.36	0.37	0.12	0.01	0.17
2	0.10	0.71	0.08	0.27	1.14	1.87	0.60	0.11	0.33	0.52	0.01
3	0.50	0.67	0.47	0.03	1.20	0.39	0.27	0.45	0.42	0.13	0.31
4	0.22	0.44	0.62	0.09	0.13	0.59	0.25	0.16	0.05	0.68	0.27
5	0.35	0.50	0.11	1.22	0.75	0.52	0.40	0.18	0.05	0.91	0.32
6	0.42	0.40	0.20	1.19	1.81	1.51	1.89	1.76	0.02	0.06	0.40
7	0.36	0.30	0.23	0.14	0.69	1.38	2.58	2.26	1.96	1.01	1.04
8	0.47	0.27	0.73	0.98	0.43	0.11	0.01	0.00	0.39	0.35	1.23
9	0.28	0.25	0.17	0.66	0.69	0.32	0.29	0.36	0.15	0.39	0.54

	Gas Cloud Number
Feature	1	2	3	4	5
Size	49	62	115	324	28
Max CL	0.74	0.99	1.83	2.69	0.93
Mean CL	0.60	0.71	0.93	1.19	0.69

Parameter	Value
Spectral range	$600 - 1800 c m^{- 1}$
Resolution for remote sensing	$4 c m^{- 1}$
Field of view with telescope	7.5 mrad
Field of regard	$360^{\circ} \times 60^{\circ}$
Scan rate ( $Δ \bar{ν} = 4 c m^{- 1}$ )	10 spectra/s
Angular resolution of the scanner	0.01°

Reconstruction of a leaking gas cloud from a passive FTIR scanning remote-sensing imaging system

Abstract

1. INTRODUCTION

2. METHODOLOGY

A. Imaging Method for the Low-Resolution CL Image of a Gas Cloud

B. Reconstruction of a High-Resolution CL Image of a Leaking Gas Cloud

C. Assessment of a High-Resolution CL Image of a Reconstructed Leaking Gas Cloud

3. SIMULATION AND EXPERIMENT

A. Passive FTIR Scanning Remote-Sensing Imaging System

B. Simulation

C. Field Experiment

4. RESULTS AND DISCUSSION

A. Simulation Results and Discussion

B. Experiment Results and Discussion

1. Imaging Process of Low-Resolution CL Images of Gas Clouds

2. High-Resolution Reconstruction of the ${{\rm{SF}}_6}$ Leaking Gas Cloud

5. CONCLUSION

Funding

Acknowledgment

Disclosures

Data Availability

REFERENCES

Data Availability

Cited By

Figures (9)

Tables (4)

Equations (13)

Applied Optics