Towards Gas Plume Identification in Industrial and Livestock Farm Environments Using Infrared Hyperspectral Imaging: A Background Modeling and Suppression Method

Abstract

Hyperspectral imaging for gas plume identification holds significant potential for applications in industrial emission control and environmental monitoring, including critical needs in livestock farm environments. Conventional pixel-by-pixel spectral identification methods primarily rely on spectral information, often overlooking the rich spatial distribution features inherent in hyperspectral images. This oversight can lead to challenges such as inaccurate identification or leakage along the edge regions of gas plumes and false positives from isolated pixels in non-gas areas. While infrared imaging for gas plumes offers a new perspective by leveraging multi-frame image variations to locate plumes, these methods typically lack spectral discriminability. To address these limitations, we draw inspiration from the multi-frame analysis framework of infrared imaging and propose a novel hyperspectral gas plume identification method based on image background modeling and suppression. Our approach begins by employing background modeling and suppression techniques to accurately detect the primary gas plume region. Subsequently, a representative spectrum is extracted from this identified plume region for precise gas identification. To further enhance the identification accuracy, especially in the challenging plume edge regions, a spatial-spectral combined judgment operator is applied as a post-processing step. The effectiveness of the method was validated through experiments using both simulated and real-world measured data from ammonia and methanol gas releases. We compare its performance against classical methods and an ablation version of our model. Results consistently demonstrate that our method effectively detects low-concentration, thin, and diffuse gas plumes, offering a more robust and accurate solution for hyperspectral gas plume identification with strong applicability to real-world industrial and livestock farm scenarios.

1. Introduction

With the increasing demands for environmental monitoring, gas identification in industrial settings and livestock farms has become critical for ensuring safety and protecting public health. Infrared spectroscopy technology [1,2,3] leverages the unique absorption or emission characteristics of different substances at specific wavelengths, capturing spectral information with distinct “fingerprint” features. This non-contact identification method holds significant potential in applications such as gas plume identification and environmental monitoring. Infrared hyperspectral imaging technology [4,5,6] further enhances this capability by simultaneously acquiring both the spatial and spectral information of a target, forming a “data cube.” This enables the precise identification of gas species based on their spectral signatures while visually representing the spatial distribution of the gases, providing robust technical support for real-time gas monitoring in complex scenarios, including industrial facilities and agricultural settings like livestock farms.

In recent years, significant advancements have been achieved in the detection and identification of gases using infrared hyperspectral imaging, driven by extensive research efforts conducted worldwide. For instance, Vallières et al. [7] proposed a robust approach for detecting, identifying, and quantifying gaseous chemicals using the Telops FIRST thermal infrared hyperspectral imager, employing a combination of clutter-matched filters (CMF) and spectral angle mapper (SAM) algorithms to achieve high sensitivity, low false alarm rates, and real-time performance. Sabbah et al. [8] employed a hyperspectral imager utilizing Fourier transform spectroscopy to measure the release of SF₆ gas near the center of Hamburg, effectively monitoring the gas plume’s diffusion. By applying a Gaussian spatial filter and a spectral correlation matching (SCM) algorithm, they successfully identified the SF₆ gas and provided real-time imagery of the gas plume’s dynamics. Omruuzun et al. [9] proposed an endmember signature-based detection method for flammable gases in longwave infrared (LWIR) detection hyperspectral images, employing the vertex component analysis (VCA) hyperspectral unmixing algorithm combined with spectral similarity metrics such as adaptive cosine estimator (ACE), SAM, matched filter (MF), and correlation coefficient to detect and quantify methyl alcohol vapor. Ozturk et al. [10] proposed a two-stage approach for CO₂ gas detection using hyperspectral images, employing SAM and correlation coefficient algorithms, where a rough detection is first performed using a narrow band and then refined using a more detailed CO₂ signature, with experiments conducted on mid-wave infrared (MWIR) hyperspectral images of vehicle exhaust. Ayhan et al. [11] proposed a systematic method for the remote detection and classification of chemical vapors using the Adaptive Infrared Imaging Spectroradiometer (AIRIS) sensor, combining adaptive background subtraction, normalization, and detection algorithms like constrained energy minimization (CEM) and SAM to effectively identify vapors such as triethyl phosphate (TEP) and dimethyl methyl phosphonate (DMMP) under varying conditions. Özdemir et al. [12] proposed a deep learning-based method for hyperspectral gas detection in the LWIR range, integrating a 3D convolutional neural network (CNN) and autoencoder-based unmixing with a neural network classifier, achieving superior detection of methane and sulfur dioxide compared to conventional methods like SAM. Schaum [13] proposed a lognormal-based gas detection algorithm for hyperspectral imaging, which optimally detects and quantifies gas plumes, such as methane, under all unknown concentrations by leveraging the Beer–Lambert law and a physically constrained background model. Despite the significant progress achieved by these methods, the majority continue to rely predominantly on the detection of spectral features, often overlooking the rich spatial distribution information inherent in hyperspectral images. This limitation can result in missed or erroneous detections along the edge regions of the gas plumes, as well as the misclassification of the discrete image elements in non-gas regions as gases, ultimately constraining the overall detection performance.

The need for robust gas identification extends to diverse environments, notably industrial settings where leaks of hazardous gases (e.g., methane, ammonia, and sulfur dioxide) pose risks to workers and the environment, and livestock farms, which are significant sources of agricultural emissions. For instance, ammonia (NH₃) is a prevalent gas in both industrial processes and livestock operations, contributing to air pollution, odor issues, and health hazards. Accurate and timely detection of such gases is crucial for safety compliance and environmental management. In livestock farms, specifically, continuous monitoring of gases like ammonia, methane, and hydrogen sulfide is vital for animal welfare, worker safety, and mitigating environmental impact [14,15,16]. However, these environments often present complex backgrounds, varying meteorological conditions, and diffuse, low-concentration gas plumes, which pose unique challenges for conventional gas detection methods.

Two-dimensional (2D) moving target detection [17,18] relies primarily on analyzing temporal changes in the image to identify moving targets. This method is widely used in the field of infrared imaging [19] for gas plume detection. Commonly employed algorithms in this domain include the optical flow algorithm [20], the inter-frame difference algorithm [21], and the background subtraction algorithm [22]. The optical flow algorithm assumes the brightness constancy of objects across consecutive frames and traces motion by analyzing pixel movement vectors. While this method achieves high detection accuracy, it is computationally intensive and sensitive to rapid target motion and variations in ambient lighting, which limits its applicability in real-time scenarios. The inter-frame difference algorithm detects targets by computing the difference between consecutive frame images, which is suitable for scenes with obvious target motion. The background subtraction algorithm addresses the detection challenges in complex environments by constructing and dynamically updating a background model. By comparing the current frame with this model, gas plume targets can be effectively separated. Among the background subtraction methods, the visual background extraction (ViBe) [23,24] algorithm is a widely adopted approach. ViBe builds a background model using multiple samples for each pixel and determines whether a pixel belongs to the plume or background by comparing its value in the current frame with those in the background model. Inspired by this, we further extend the idea of background subtraction to hyperspectral image processing, drawing an analogy between the spectral dimension and “timeline”. When the target gases exhibit distinct absorption or emission characteristics in specific spectral bands, the images corresponding to these characteristic bands display significant differences. Such image differences can be leveraged to identify target gases effectively, offering novel insights for the application of hyperspectral imaging technology in gas plume detection.

Therefore, the primary purpose of this study is to develop and validate a novel identification method that overcomes the limitations of traditional pixel-wise approaches by effectively leveraging spatial information. The key objectives are to reduce false positives caused by spectral similarity alone and to improve the accuracy of plume boundary detection, especially in challenging low-concentration regions. To achieve these objectives, the main contributions and innovations of this paper are as follows: First, we introduce a new conceptual framework by adapting the principle of dynamic background modeling and subtraction from temporal video analysis to the spectral domain, creating a robust spatial-first approach to locate the plume. Second, we designed a spatial-spectral combined judgment operator that synergistically integrates spatial variability and spectral correlation to significantly enhance detection precision at diffuse plume edges. Finally, the effectiveness and superiority of the proposed method are validated through experiments on both simulated and real-world measured data, indicating its strong application potential for gas plume identification in complex environments like industrial and livestock farm settings.

2. Methods

The key steps of the proposed method are elaborated as follows: First, a background model was constructed, followed by band-by-band plume detection to suppress the background and detect the gas region. Representative spectral analysis was performed to accurately classify the gas type. Finally, edge post-processing was applied to refine gas boundary detection. The flowchart of the proposed method is shown in Figure 1.

2.1. Image Background Modeling and Suppression

In a hyperspectral image, each band corresponds to a distinct image. By drawing inspiration from foreground detection techniques in image processing and extending these methods to hyperspectral image analysis, it becomes possible to analyze each spectral band image, thereby effectively identifying the regions of gas plumes.

In constructing the background model, our approach is inspired by the initialization principle of the foundational ViBe [23,25] algorithm, which leverages spatial information to build a model from a single frame. We adapted and extended this principle for hyperspectral imagery. Considering that infrared images often have lower resolution and less distinct texture than visible-light images, relying on a single spectral band may not provide sufficient sample diversity for a robust initial model.

Therefore, to enhance diversity, we selected a small set of initial bands for background modeling. As a common characteristic of hyperspectral imaging systems, which often span hundreds of contiguous spectral bands, the channels at the very beginning of the instrument’s operational range are typically less likely to contain the primary absorption features of many target gases. Based on this principle, we chose the first three spectral bands as our initial background sources. This choice was then validated for our specific experiments; for the hyperspectral imaging spectrometer used, the first three bands were located at the short-wavelength edge of the instrument’s range and, according to the NIST (National Institute of Standards and Technology) reference spectra, contained no measurable absorption features for both ammonia and methanol. We note that this selection criterion is instrument-dependent and gas-specific; for other sensors or target gases, the appropriate non-absorption bands should be determined based on the instrument’s spectral coverage and the target’s known spectral characteristics.

To further enrich the model, we then introduced a spatial neighborhood random sampling strategy. For each of the three initial single-band images, two additional images were generated by replacing each pixel’s value with one randomly selected from its 8-connected neighborhood. By applying this operation to each of the three initial bands, the original 3 images were expanded to a set of 9 images. Thus, the background model M(x) for a pixel at position x consisted of the N = 9 pixel values from this expanded set. It should be noted that while ‘three’ serves as a practical choice for the number of initial bands in this study, this parameter can be adapted based on the specific characteristics of the sensor and the scene. The background sample M(x) at position x can be expressed as follows:

$$M\left(x\right)=\left\{p_1,p_2,\cdots ,p_N\right\}$$

(1)

For a hyperspectral image with dimensions H × W × B (where H denotes the image height, W denotes the image width, and B denotes the number of bands), the constructed background model comprises W × H × N sample values. Here, M(x) is a three-dimensional vector with dimensions H × W × N, which can be understood as a collection of N images. The process of constructing the background sample model is shown below (Figure 2).

After constructing the background model, plume detection was performed on the images of each band. Let the pixel value of a pixel x in the image of the n-th band (n > 3) be p_n(x).

For pixel x, we compute the absolute difference against all corresponding sample points in M(x) and sum these differences to obtain Δ_s, which serves as a quantitative measure of the pixel’s deviation from the background model. The scalar quantity Δ_s undergoes continuous updates in response to band changes. The equation for Δ_s is given as follows:

$${\Delta }_s = {\displaystyle \sum _{i=1}^N\left|p_n(x)-p_i\right|}$$

(2)

where p_n(x) represents the pixel value at spatial position x within the n-th band, and p₁, p₂, …, p_N are the corresponding sample points in the background model M(x).

Define a circular region S_R of radius R centered at pixel x. If the number of background samples within this circular region is greater than or equal to the preset threshold D, pixel x is classified as a background pixel; otherwise, it is classified as a plume pixel, as illustrated in Figure 3.

To adapt to changes in the environment and background, it is necessary to update the background model after the detection process. This is achieved by randomly replacing a pixel value in the background model of a pixel detected as non-plume with a corresponding pixel value from the current frame. This approach dynamically incorporates the latest scene information into the background model. Once the current band image is processed, the next band image is analyzed, enabling plume detection for each band image in the hyperspectral dataset. Since the target gas typically exhibits absorption or emission peaks, the plume regions in the images are more pronounced when the selected band is located near the strongest peak region. We established a gas spectral database including precise information on the positions of characteristic absorption peaks for various gases. The standard reference spectra for target gases were sourced from the NIST Chemistry WebBook. The NIST database provides high-resolution transmittance spectra, which must be adapted to be comparable with the lower-resolution data acquired by our hyperspectral instrument.

To achieve this, a spectral resampling process was performed. For each target gas, its high-resolution spectrum from NIST was interpolated to match the exact band centers and spectral resolution of our experimental data. For example, for the simulated and measured ammonia datasets, the reference spectrum was resampled to 162 bands spanning 870 cm⁻¹ to 1250 cm⁻¹. This process ensures that the reference spectrum and the measured spectra are directly comparable on a band-by-band basis. The adapted, instrument-specific reference spectrum was then used for two purposes: (1) to guide the initial plume detection (Section 2.1), where the result from the single band corresponding to the spectrum’s strongest absorption feature is ultimately used to determine the primary plume region; and (2) its resampled spectral vector was used for the correlation-based edge refinement (Section 2.2).

To enhance both the efficiency and accuracy of gas plume detection and identification, we implemented a staged approach. In the preliminary detection phase, we focused on the strongest absorption band of the target gas, leveraging its prominent absorption features to effectively localize the plume region spatially. This approach also reduces computational complexity. However, relying on a single band for detection presents certain limitations, particularly in terms of potential false alarms or omissions, especially at the plume edges. This initial detection determines the primary spatial distribution of the target gas, which is then refined in a subsequent post-processing stage to improve edge accuracy.

2.2. Post-Processing of Edge Regions

After detecting the plume region, further analysis is required to determine whether it corresponds to the target gas. Pixel spectra detected as belonging to the plume region need further processing to produce a representative spectrum. This is achieved through one of the following approaches: principal component analysis (PCA), which involves selecting the primary eigenvectors of the plume pixel spectra to reconstruct a representative spectrum and effectively suppress noise, and the averaged spectral method, which calculates the averaged spectrum of the plume region as the basis for subsequent discrimination. We employed the averaged spectral method. This approach generates representative spectra by averaging the pixel spectra within the plume region, which serves as the foundation for subsequent identification. We selected the averaged spectral method primarily due to its computational efficiency. Moreover, in this study, the approximate boundaries of the gas plume were predefined, and the spectral features within the plume region were relatively homogeneous. As such, the averaged spectral method effectively captures the overall spectral characteristics of the plume region. In contrast, PCA reconstructs representative spectra by selecting the principal eigenvectors, which helps to effectively suppress noise and is typically employed in cases where spectral data are heavily contaminated by noise. The resulting representative spectrum is then compared with a standard spectrum or pre-established gas spectral libraries using similarity metrics such as the spectral angular distance or correlation coefficient. If the similarity exceeds a predefined threshold, the target gas class within the plume region can be reliably detected.

The edge of a gas plume is typically characterized by its thinness and dispersion in space, resulting in weakened spectral features. This makes traditional approaches, such as plume detection or spectral similarity discrimination, prone to omissions and misdetections in the edge regions of the plume. To address this challenge, this study builds upon existing detection and category identification methods while considering the unique characteristics of the plume edge. Recognizing that reliance on a single index is susceptible to interference, we propose a spatial-spectral integrated discrimination operator. The operator is defined by the following equation:

$$\begin{array}{ll}{C}_{sp}& =\alpha \times C_s+\beta \times C_{\lambda }\\ & =\alpha \times {\displaystyle \frac{{\Delta }_s}{R\times D}}+\beta \times {\displaystyle \frac{r_{\lambda }}{T_{\lambda }}}\end{array}$$

(3)

Here, Δ_s represents the sum of the absolute differences between the pixel values of the current band image and the background model. When the pixel values of the current band image are close to the background model, Δ_s is small, suggesting that the region is unlikely to exhibit gas absorption features. Conversely, when the pixel values deviate significantly from the background model, Δ_s increases, indicating a higher likelihood of the presence of gas absorption features. Therefore, Δ_s serves as an indicator of spatial image variability, with larger values reflecting greater discrepancies between the image and the background model, and a higher probability of gas presence. r_λ denotes the correlation coefficient between the spectra of neighboring pixels and the standard spectrum, indicating their spectral similarity. α and β are weight coefficients that control the relative influence of spatial variability and spectral correlation within the combined judgment operator, and α + β = 1. R denotes the radius of the circular region S_R centered at pixel x, which is used to evaluate the spatial consistency of local background samples. D is a threshold indicating the minimum number of background pixels required within S_R; it helps determine whether a pixel is classified as background or plume, as previously defined after Equation (2). T_λ serves as a threshold for spectral correlation to normalize the correlation coefficient. When C_s > 1, the plume region can be approximately identified using the image foreground detection method. Similarly, when C_λ > 1, the spectral judgment operator identifies the point as a target pixel. The spectral similarity component, C_λ, is crucial for this refinement. The rationale for using spectral data here is that while the initial detection relies on the strongest absorption band for efficiency, the edge regions consist of mixed pixels where this single peak may be weak or noisy. However, the overall shape of the spectrum may still retain the characteristic signature of the gas, providing a more robust basis for identification. The mechanism for calculating this similarity is as follows: for each pixel being evaluated in the expanded edge region, its spectral vector is correlated against the standard spectrum signature of the target gas (obtained from the NIST database). We employed the Pearson correlation coefficient (r_λ) for this purpose. This process yields a scalar similarity score that quantifies how closely the pixel’s spectral “fingerprint” matches that of the target gas, ensuring a reliable refinement of the plume boundary. Our operator integrates both image-based and spectral-based detection methods. The contour plot of the combined judgment operator is presented in Figure 4, using α = β = 0.5 and C_sp = 0.8 as an example. In this coordinate plane, regions with horizontal coordinates greater than 1 represent pixels identified as targets through plume detection, while regions with vertical coordinates greater than 1 correspond to pixels identified as targets through spectral correlation identification. The upper-right corner region of the contour line (here 0.8) characterizes that the combined judgment operator ultimately judged them as targets.

In this study, a neighborhood expansion technique is employed to process the edges. Specifically, the method centers on the initially detected edge points and expands them within a 5 × 5 pixel neighborhood to extend the edge line into the edge region, thus enabling more effective detection over a larger area. Subsequently, the expanded edge region is analyzed using a comprehensive discriminant operator. The threshold value for this operator is determined experimentally for each target gas species. Through repeated experiments, the optimal threshold parameter that best distinguishes the gas plume from the background is selected, taking into account the spectral characteristics of different gases. If the value of the integrated discriminant operator exceeds the predefined threshold, the pixel is classified as part of the gas plume; otherwise, it is regarded as background or noise. The detection process continues iteratively until the overlap between the foreground regions detected in successive iterations exceeds a preset threshold of 80% [26,27], at which point the edge detection process is considered complete. This post-processing step is designed to enhance the algorithm’s sensitivity to weak edge features while minimizing the accumulation of errors from pseudo-edge information during the iteration process.

The time complexity of our method reflects a strategic trade-off. Its initial detection stage has a complexity of O(B_peak × H × W × N), where H and W are the image dimensions, N is the number of background samples, and B_peak is the band index of the target’s strongest absorption peak. This stage is theoretically faster on average than a full-cube analysis because it often terminates early (B_peak < B, where B is the total number of bands), significantly reducing the computational load. The algorithm then adds a computationally intensive post-processing stage with a complexity of O(I_iter × P_edge × K_hood × B), where I_iter is the number of iterations, P_edge is the number of detected edge pixels, and K_hood is the neighborhood size for refinement. Although this second stage involves spectrum operations, its computational cost is localized to a very small subset of pixels (P_edge). By minimizing computation in the initial stage and concentrating resources on a targeted, high-precision refinement of edge pixels, the method achieves an effective balance between overall computational efficiency and localized accuracy.

2.3. Experimental Datasets

To evaluate the effectiveness of the proposed method, a set of simulated hyperspectral images is first used in this study for experiments. The advantage of using simulated data lies in its ability to provide explicit ground truth, which refers to known information about the actual target, an essential component for evaluating the accuracy and performance of the algorithm. By embedding an ammonia plume into an existing hyperspectral image background, a simulated hyperspectral image containing ammonia is generated, allowing for precise control over the target features. The algorithm’s performance in target detection and identification is then comprehensively evaluated by comparing the results with the known ground truth. Additionally, two sets of constrained hyperspectral images are employed in this study to further validate the algorithm. The use of real-world data enhances the evaluation by testing the algorithm’s effectiveness and robustness in practical scenarios, thus ensuring its reliable performance in complex environments.

Unlike traditional image simulation and spectral simulation methods, hyperspectral image simulation requires not only the simulation of the 2D gas plume distribution but also the consideration of the spectral characteristics of the plume region. Since the background temperature variation at different locations significantly impacts the simulated spectra, we selected a 100 × 100-pixel region in the lower left corner of the field of view, where the temperature difference is relatively small, as the background. This was done within a 320 × 256-pixel hyperspectral image with 162 bands in the spectral range of 870 cm⁻¹ to 1250 cm⁻¹, and the corresponding hyperspectral image was acquired. The background averaged images are shown in Figure 5, where the left map represents the averaged image of the hyperspectral image (320 × 256 pixels), and the right map shows the averaged image of the background extracted from the 100 × 100 pixels in the lower left corner. The averaged image is solely intended to illustrate the composition of the scene. The final image will overlay the gas plume detection results on this background to present the complete visualization. Figure 5 shows an averaged visual representation of the selected background scene. In the subsequent simulation process, the underlying hyperspectral data from this region served as the radiometric foundation, upon which a synthetic gas plume was superimposed to generate the test data.

In the process of generating gas hyperspectral images, a simulation framework was designed based on gas-free hyperspectral images and standard gas spectra to produce hyperspectral images containing gas plumes of various shapes. First, a 2D plume concentration distribution map was generated using computational fluid dynamics (CFD) [28,29] software. The 2D plume concentration distribution map, along with its schematic overlay on the background averaged image, is shown in Figure 6. The concentration levels are visualized through the chromatic scale, where warmer hues indicate higher molar concentrations.

On this basis, the spectra of the plume regions were simulated using the standard spectra of the gas and the plume concentration distribution, while the spectra of the non-plume regions were derived from the background spectra at their respective positions. Finally, the simulation of the infrared hyperspectral image of the gas was completed by arranging the spectra of both the plume and background regions to form the complete hyperspectral image (data cube).

For infrared spectral simulations in the plume region, the process can be modeled using a passive infrared telemetry three-layer radiative transfer model [30,31,32], as illustrated in Figure 7.

In this model, layer 1 represents the atmosphere between the gas cloud and the spectrometer, layer 2 corresponds to the target gas cloud, and layer 3 is the background.

Let L denote the radiance received by the spectrometer, and ${\tau }_i$ represent the transmittance of layer $i\left(i=1,2\right)$. Let $B_i$ denote the blackbody radiance at the temperature of layer i, and L₃ denote the background radiance. The radiance at the instrument’s entrance, considering the presence of clouds, is given as follows:

$$L=(1-{\tau }_1)B_1+{\tau }_1[(1-{\tau }_2)B_2+{\tau }_2{L}_3]$$

(4)

where ${\tau }_2=e^{-\alpha CL}$ denotes the gas cloud transmittance, α represents the absorption coefficient of the gas cloud, C represents the concentration, and L denotes the effective path length. When combined, the product CL refers to the column density.

The radiance at the instrument’s entrance in the absence of gas clouds can be expressed as follows:

$$L_0=(1-{\tau }_1)B_1+{\tau }_1{L}_3$$

(5)

It is generally assumed that the first atmospheric layer is thin, and its transmittance ${\tau }_1$ is close to 1, such that $L_0\approx L_3$. As the gas cloud diffuses, it exchanges heat with the atmosphere, causing the gas cloud temperature to rapidly approach the atmospheric temperature. Consequently, the gas cloud temperature can be approximated as the atmospheric temperature, and thus $B_1\approx B_2$. Therefore, in the presence of a gas cloud, the radiance at the instrument entrance can be expressed as follows:

$$L=(1-{\tau }_2)B_1+{\tau }_2{L}_0$$

(6)

Based on the above equation, different synthesized gas spectra can be obtained by varying the CL values and environment temperatures for specific target gases. Figure 8 shows an example of this synthesis process. Figure 8a displays a measured background spectrum (corresponding to the term L₀ in the model). In contrast, Figure 8b illustrates the resulting simulated spectrum (L) after the forward model Equation (6) is applied to account for the spectral effects of an ammonia plume. This synthesis approach allows the algorithm’s performance to be evaluated under precisely controlled conditions.

Additionally, two sets of constrained hyperspectral images were employed in this study to further validate the algorithm. The use of real-world data enhances the evaluation by testing the algorithm’s effectiveness and robustness in practical scenarios, thus ensuring its reliable performance in complex environments. Measured dataset 1 consists of a hyperspectral image from a scene containing ammonia gas. The data have a resolution of 320 × 256 pixels and span 162 spectral bands, ranging from 870 cm⁻¹ to 1250 cm⁻¹ [33]. An averaged image of the hyperspectral cube is presented in Figure 9a, depicting the scene of this real-world experiment. Measured dataset 2, acquired using the Telops Hyper-Cam, contains methanol gas with a resolution of 128 × 128 pixels and a spectral range of 165 bands, covering 867 cm⁻¹ to 1288 cm⁻¹ [34]. The averaged images of both hyperspectral datasets are presented in Figure 9b.

Previous studies on passive infrared standoff detection have reported typical operational parameters for successful gas identification. For example, gas plumes have been reliably detected at distances ranging from several meters to the kilometer scale, including approximately 1.3–3 km in industrial and urban monitoring scenarios [8,35]. Stable detection performance has been documented under moderate wind conditions (about 3–6 m/s) and with various gas release rates (e.g., 8–10 g/s) [36]. These works also emphasize the importance of a sufficient thermal contrast between the plume and its background for maintaining signal detectability.

The literature consistently shows that these parameters critically influence detection performance. Thermal contrast is the dominant factor controlling signal strength; wind speed and direction determine plume geometry and dispersion; and the release rate directly affects the in-plume gas concentration. Deviations in any of these parameters can substantially alter the detectability of the target gas.

In our measured datasets, exact values for parameters such as sensor-to-plume distance, meteorological conditions, and release rates were not systematically obtained. Nevertheless, the measurements were conducted within the operational envelope reported in prior studies, ensuring that the acquired data remain representative for testing algorithmic robustness under realistic and complex monitoring conditions.

2.4. Evaluation Metrics

To quantitatively assess the performance of the algorithms, we used several metrics: detection rate (DR), false alarm rate (FAR), F₁ score, and intersection over union (IoU).

The DR represents the ratio of the number of pixel points correctly identified as targets to the total number of actual target pixels. A higher DR indicates better algorithm performance. The DR is calculated as follows:

$$\mathrm{DR}={\displaystyle \frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FN}}}$$

(7)

The FAR represents the ratio of pixels incorrectly identified as targets to the total number of pixels in the actual negative samples. A lower FAR indicates better algorithm performance. The FAR is calculated as follows:

$$\mathrm{FAR}={\displaystyle \frac{\mathrm{FP}}{\mathrm{FP}+\mathrm{TN}}}$$

(8)

The F₁ score is a comprehensive evaluation index of the classification model, which is the harmonic average of precision and recall. A higher F₁ score indicates better algorithm performance. The F₁ score is expressed by the following equation:

$${\mathrm{F}}_1={\displaystyle \frac{2\times \mathrm{Precision}}{\mathrm{Precision}+\mathrm{Recall}}}$$

(9)

Let the predicted target mask be denoted as P and the ground truth mask as G. The IoU is defined as the ratio of the intersection to the union of these two masks. A higher IoU value indicates better algorithm performance. The IoU is calculated as follows:

$$\mathrm{IoU}={\displaystyle \frac{|\mathrm{P}\cap \mathrm{G}|}{|\mathrm{P}\cup \mathrm{G}|}}$$

(10)

3. Results and Discussion

3.1. Experimental Results on Synthetic Datasets

Gas plume identification was conducted on the simulated hyperspectral images, with results compared across SCM, SAM, sparse representation (SR) [37], and the proposed approach. SCM measures spectral similarity by computing the correlation coefficient between target and reference spectra, while SAM evaluates the spectral angle to reduce sensitivity to illumination variations. SR models each pixel spectrum as a sparse linear combination of reference spectra from a dictionary, with detection based on reconstruction residuals. An ablation experiment was conducted on the proposed approach by omitting the edge post-processing step, and the results were compared with SCM, SAM, SR, and the complete version. For a fair comparison, the detection threshold was selected as the one yielding the highest F₁ score. This approach ensured that all methods were evaluated at their respective optimal operating points, thereby avoiding performance bias caused by suboptimal parameter choices.

The detection outcomes for ammonia, overlaid on the background-averaged image, are shown in Figure 10; the ground truth and detection result masks for all three methods are presented in Figure 11; and the comparison of true positives (TP), false positives (FP), true negatives (TN), and false negatives (FN) is illustrated in Figure 12.

Due to the challenges of accurately measuring gas concentration in infrared spectroscopy and hyperspectral imaging, we utilized relative concentration values to indicate changes in gas concentration. The color bars in Figure 10 (and in other corresponding figures) reflect these relative values. Similarly, for methanol, the detection outcomes are shown in Figure 13; the corresponding ground truth and detection result masks are presented in Figure 14; and the comparison of TP, FP, TN, and FN is illustrated in Figure 15.

The performance histograms for ammonia and methanol detection are presented in Figure 16 and Figure 17, with the corresponding quantitative results summarized in

Table 1. Comparison of ammonia detection performance.

Methods	FAR	DR	F1 Score	IoU
SCM	0.0276	0.7554	0.8275	0.7058
SAM	0.0283	0.8012	0.8554	0.7474
SR	0.0480	0.8112	0.8390	0.7226
Ablation	0.0075	0.8087	0.8848	0.7935
Proposed	0.0337	0.9642	0.9406	0.8879

and

Table 2. Comparison of methanol detection performance.

Methods	FAR	DR	F1 Score	IoU
SCM	0.0230	0.7193	0.8041	0.6723
SAM	0.0694	0.7520	0.7663	0.6211
SR	0.0541	0.7528	0.7853	0.6466
Ablation	0.0097	0.8502	0.9046	0.8258
Proposed	0.0045	0.8647	0.9207	0.8530

. For both gases, the proposed method achieves the highest DR, F₁ score, and IoU among all compared algorithms, while keeping the FAR at a relatively low level. Specifically, for ammonia detection, the proposed method attains a DR of 0.9642, an F₁ score of 0.9406, and an IoU of 0.8879, surpassing the other methods by a considerable margin. For methanol detection, it similarly yields a DR of 0.8647, an F1 score of 0.9207, and an IoU of 0.8530, while maintaining the lowest FAR (0.0045). These results demonstrate the robustness and generalization capability of the proposed method across different gas types, providing more accurate and reliable detection performance than SAM, SCM, SR, and the ablation model.

From the quantitative results in Table 1 and Table 2, the specific sources of error for each method can be identified. The primary error source for SCM, SAM, and SR is their pixel-wise nature, which completely disregards spatial context. For instance, in the ammonia test (Table 1), while SAM achieves a decent DR of 0.8012, its IoU is only 0.7474. This discrepancy indicates that while it finds the plume’s core, it either misses significant portions of the diffuse edges (contributing to false negatives) or, as suggested by its FAR, it incorrectly identifies many isolated background pixels as gas (false positives). This leads to a poorly defined plume shape and a low IoU.

The “ablation” model’s results precisely pinpoint the importance of our edge post-processing step. In both the ammonia and methanol tests, the ablation model achieves a very low FAR (0.0075 and 0.0097, respectively), demonstrating that our background modeling stage is highly effective at suppressing false positives. However, its DR and IoU are consistently lower than the final proposed method. This shows that its main source of error is the inability to accurately capture the weak signals at the plume’s edges, leading to incomplete boundary detection.

In contrast, the proposed method systematically mitigates these errors. Its initial background modeling stage overcomes the false positive issue inherent in pixel-wise methods, and its subsequent edge post-processing module specifically corrects the boundary inaccuracies that the ablation model suffers from. This two-stage, spatial-spectral synergy is why it consistently achieves the highest IoU (0.8879 for ammonia, 0.8530 for methanol), reflecting a more complete identified plume.

3.2. Experimental Result on Measured Datasets

In the previous experiments, the effectiveness of the proposed algorithm was validated using simulated data. To further assess its performance in practical applications, this section employs measured data to verify and analyze the algorithm’s performance.

The experimental results for the ammonia and methanol hyperspectral data, processed using the SCM, SAM, SR, ablation, and proposed methods, are shown in Figure 18 and Figure 19, respectively. The ammonia detection results in Figure 18 are overlaid on the experimental scene from Figure 9a, while the methanol results in Figure 19 are overlaid on the corresponding experimental scene from Figure 9b. This enables a direct visual comparison of their performance in capturing the gas distribution.

In Figure 18 and Figure 19, the background regions are displayed in grayscale, while the detected gas regions are shown in pseudo-color, with warmer colors indicating higher relative concentrations. For both gases, the SCM and SAM methods capture the majority of gaseous regions but produce relatively coarse results, with more scattered pixels identified as gas pixels. Spectral analysis reveals that these scattered points lack the characteristic absorption features of the target gas species, indicating false positives. The SR method offers slightly improved continuity of detected regions but still exhibits noticeable false detections.

For ammonia detection (Figure 18), the proposed method accurately identifies nearly all gaseous regions, yielding a more continuous and complete representation of the plume and aligning well with the natural diffusion pattern. For methanol detection (Figure 19), the superiority of the proposed method is also evident: it effectively suppresses false positives, maintains plume integrity, and captures fine structural details of the gas distribution. However, it is worth noting that a detailed inspection reveals some minor missed detections (false negatives) in the most diffuse, low-concentration area at the top-left of the plume.

Regarding the evaluation of the measured datasets, we primarily adopted a qualitative visual assessment. We recognize that this is mainly due to the fundamental challenge of generating a precise, pixel-level “ground truth” for dynamic and diffuse real-world gas plumes. The constantly changing and irregular nature of these plumes makes the direct application of quantitative metrics like IoU very difficult, which is a common challenge in the field of standoff gas detection. Therefore, we acknowledge that the absence of quantitative metrics for this part of the study is a limitation. We chose to rely on visual evaluation against the original imagery to assess the physical plausibility of the detection results, which serves as a feasible supplementary validation under these constraints.

Despite the limitation, across both gases, the proposed method provides the highest detection accuracy and most faithful visualization of gas plumes, demonstrating strong generalization capability to different gas types. The primary advantage of our approach lies in its novel spatial-first framework, which adapts background modeling from temporal analysis to the spectral domain. By first identifying the main plume region spatially, it effectively mitigates the scattered false positives that plague traditional pixel-wise methods like SCM and SAM. Furthermore, the dedicated spatial-spectral operator provides a significant advantage in accurately delineating diffuse plume edges, a common failure point for other methods, leading to a more complete and realistically shaped plume detection as evidenced by the superior IoU and F₁ scores.

The method has several limitations. First, its initial plume localization is predicated on the existence of a distinct and strong characteristic absorption peak, which may compromise performance for gases with very weak or obscured spectral features. Furthermore, as observed in the experimental results, there is an inherent trade-off between its excellent false positive suppression and its sensitivity at the most diffuse, low-concentration plume edges, where minor missed detections can occur. Finally, the method’s reliance on several experimentally determined parameters suggests that some degree of manual tuning may be required when applying it to new targets or significantly different environments. Addressing these limitations through more adaptive and automated strategies will be the focus of our future work.

4. Conclusions

In this study, we developed and validated a novel spatial-first identification method for infrared hyperspectral gas plumes, designed to overcome the limitations of traditional pixel-wise spectral methods by adapting the principle of background modeling from the temporal to the spectral domain. This two-stage design first integrates spatial information to significantly reduce false positives and then employs a spatial-spectral post-processing framework to improve edge-region accuracy, proving especially effective in low-concentration and diffuse plume conditions. Experimental evaluations on both simulated and measured datasets—including ammonia and methanol gas—demonstrate that the proposed approach consistently outperforms conventional methods. These results confirm the method’s robustness and applicability to complex real-world scenarios, such as industrial emission monitoring and livestock farm environmental management. Future work will explore the integration of real-time processing, multi-gas simultaneous detection, and deployment on airborne or stationary platforms to further extend the practical utility of the approach.