EP1430280A2

EP1430280A2 - Small spectral signal detection system

Info

Publication number: EP1430280A2
Application number: EP02803278A
Authority: EP
Inventors: Kwong W. Au; Darryl G. Busch; Saad J. Bedros; Rida M. Honeywell International Inc. HAMZA; Jess E. Fordyce
Original assignee: Honeywell International Inc
Current assignee: Honeywell International Inc
Priority date: 2001-09-24
Filing date: 2002-09-24
Publication date: 2004-06-23
Also published as: WO2003054824A2; AU2002365125A1; WO2003054824A3; WO2003054824B1; AU2002365125A8

Abstract

A standoff chemical agent detection system passively detects chemical agents in a scene. Interferograms are generated from received scene spectral information. In one embodiment, the spectral information is apodized, and a chirp Fast Fourier Transform is performed on the interferogram. A calibration curve is applied to correct gain and offset, and a feature vector is generated based on comparison of the normalized spectrum to predetermined chemical shape templates. The feature vector is provided to a classifier to identify the existence and identity of a chemical agent threat. The chemical agent detection system utilizes a lower resolution search mode, and a higher resolution confirmation mode, wherein both modes utilize the same detection algorithms. The neural network is trained using a large database of training data. The network is iteratively trained using partially random subsets of the training data. Problematic data from previous subsets is included in further subsets to improve the training.

Description

Small Spectral Signal Detection System

This application claims the benefit of priority under 35 U.S.C 119(e) to U.S. Provisional Patent Application Serial No. 60/324397, filed September 24, 2001 , the entirety of which is incorporated herein by reference.

Government Funding

The invention described herein was made with U.S. Government support under subcontract number: LS97-00001 under Prime Contract Number DAAM01-97- C-0030 awarded by U.S. Army, SBCCOM, Edgewood, MD. The United States Government has certain rights in the invention.

Field of the Invention

The present invention relates to identification of small signals embedded in a large background signal, and in particular to an algorith for detecting chemical vapors such as, chemical agents and toxic industrial chemicals.

Background of the Invention

The threat of attack on military and civilian targets employing chemical warfare agents is of growing concern. Various technologies are being developed for the detection and identification of such agents. Standoff chemical agent detectors are needed for real time, on the move detection for contamination avoidance and reconnaissance operations on a wide variety of land, air, space and sea platforms.

Several technical challenges need to be solved for the use of passive systems, e.g., Fourier Transform Infrared (FTIR) spectrometer, and hyper-spectral spectrometer to detect agent clouds at a distance using only thermal emission from the scene. Agent emission/absorption signatures are received against an unknown and constantly changing background. These types of backgrounds are encountered when the detectors are located on a moving platform. Reference backgrounds from previous scans are not reliable with moving platforms. Further, large areas need to be searched, requiring high processing power to obtain confirmation of detected threats. High false alarm rates occur when trying to detect low agent signature strengths.

Summary of the Invention A standoff chemical agent detection system passively detects chemical agents in a scene. Interferograms are generated from received scene spectral information. In one embodiment, the interferogram is apodized, and a chirp Fast Fourier Transform is performed on the apodized interferogram. A calibration curve is applied to correct system gain and/or offset, and a feature vector is generated based on comparison of the transformed spectral information to predetermined chemical agent shape templates. In one embodiment, the feature vector is provided to a classifier to identify the existence and identity of a chemical agent threat. hi one embodiment, the system calibration is a function of one or two known temperature sources, generating a gain and/or an offset to apply to the spectra. Apodization is performed to correct for the effect of asymmetry on the interferogram. The spectral transform is a chirp type of transform that allows sampling of data at a selected frequency comb to calibrate between laser reference frequencies. In another embodiment of this present invention, a zoom Fast Fourier Transform can be used rather than the Chirp Fourier Transform to obtain the spectral signature. The system gain and or offset is applied to the spectra followed by a normalization with a

Planck's function. The characteristics of the normalized spectra are quantified by a feature vector, which consists of a set of features. Each feature is the results of a comparison of a chemical vapor shape template to the selected normalized spectral region using a least squares fit algorithm. For each shape template, values for the amplitude, slope, offset, and mean square error are calculated. Characteristics of each potential agent are captured in a specific subset of the feature vector. The feature vector subset for each threat is processed by a separate neural network for the detection of each potential threat. In another embodiment of the present invention, the vector subsets are processed by one neural network with multiple layers that is trained to process all vectors simultaneously. The presence or absence of the chemical threat is made based on a sequential decision making process wherein processing is stopped once a decision is made regardless of the existence of further data to process. In a further embodiment, the chemical agent detection system utilizes a lower resolution search mode, and a higher resolution confirmation mode, wherein both modes utilize the same detection algorithms with different shape templates and neural network coefficients. The neural network is trained using a large database of training data. The network is iteratively trained using partially random subsets of the training data. Problematic data from previous subsets is included in further subsets to improve the training.

Brief Description of the Drawings FIG. 1 is a block diagram of a passive mobile chemical agent detection system. FIG. 2 is a high level block flow diagram of the chemical agent detection system of FIG. 1. FIG. 3 is a high level block flow diagram of algorithms used to process interferograms generated by the chemical agent detection system of

FIG. 1. FIG. 4 is a block flow diagram of preprocessing of an interferogram.

FIG. 5 is a block flow diagram of an algorithm for extracting a feature vector from a normalized spectrum FIG.s 6A, 6B, 6C and 6D are representations of multiple different shape templates used to represent known agent peaks and common interferents. FIG. 7 is a block flow diagram of classifying feature vectors to identify chemical agents.

Detailed Description of the Invention

In the following description, reference is made to the accompanying drawings that form a part hereof, and in which is shown by way of illustration specific embodiments in which the invention may be practiced. These embodiments are described in sufficient detail to enable those skilled in the art to practice the invention, and it is to be understood that other embodiments may be utilized and that structural, logical and electrical changes may be made without departing from the scope of the present invention. The following description is, therefore, not to be taken in a limited sense, and the scope of the present invention is defined by the appended claims. The functions or algorithms described herein are implemented in software in one embodiment, where the software comprises computer executable instructions stored on computer readable media such as memory or other type of storage devices. The term "computer readable media" is also used to represent carrier waves on which the software is transmitted. Further, such functions correspond to modules, which are software, hardware, firmware of any combination thereof. Multiple functions are performed in one or more modules as desired, and the embodiments described are merely examples. The software is executed on a digital signal processor, ASIC, microprocessor, or other type of processor operating on a computer system, such as a personal computer, server or other computer system.

A chemical agent detection system for use in detecting chemical agent clouds in a mobile setting is shown generally at 100 in FIG. 1. The system is housed in an enclosure 110 in one embodiment, and is mounted on a moving platform 120, such as a moving vehicle, whether ground, water space or air based. The platform can also be stationary at a fix site. The chemical agent detection system 100 is used to detect and differentiate chemical agent vapors 125 by class and by type with a very low false alarm rate. To meet this objective, a large field of regard is interrogated within defined time constraints on many application platforms under numerous conditions. The chemical agents to be detected include three classes: nerve, blister, and blood; each class has many agent species. The numerous conditions include ideal and real battlespace environments, with or without common battlefield interferents, and viewing various types of backgrounds 130. The many applications include sea, land, space or air operation while stationary or on a moving platform. The diversity of operational environments imposes many challenges to the objective. The key challenges include real time processing, sensor ruggedization, wide area spatial scanning, detection/discrimination without a previously known background reference, and size and weight limitations.

One type of chemical agent detection system utilized employs passive sensing of infrared (LR) emissions. The emissions, along with background emissions are received through a window 132 mounted in the enclosure 110, and focused by a second lens system 136 onto a beam splitter element 140. Some of the IR is transmitted by a first stationary mirror 144 mounted behind the beam splitter element 140. The rest of the IR is reflected by element 140 onto a moving mirror 146. The reflected beams from the stationary mirror 144 and moving mirror 146 combine to create an interference pattern, which is detected by an IR detector 148. An output of the IR detector is sampled in one of two modes to create an interferogram, which is processed at a processor 160 to provide an output 170 such as a decision regarding whether or not a threat exists. In a search mode as indicated at 210 in FIG. 2, a reduced resolution is utilized at approximate a 16 wavenumber resolution. When potential agents are detected, the mode is switched at 220 to a confirmation mode with sequential decision making at 230. At 240, the extent of the potential threat is mapped to provide an indication of the size and location of the threat. Algorithms are utilized to detect chemical agents as shown in FIG. 3. The algorithms use the reduced resolution (16 wavenumber)

"search mode", but without loss of detection sensitivity relative to the 4 wavenumber resolution "confirm mode". Four wavenumber resolution has been used in detecting chemical agents with great success. However, the time to scan the entire field of range at 4 wavenumber resolution would exceed time constraints in many applications and would not provide enough time to take protective measures or to take evasive action for contamination avoidance. The time to acquire radiometrically equivalent 16 wavenumber resolution data is 16 times less than that for 4 wavenumber resolution data. The 16 wavenumber resolution data does not provide as much detail as the 4 wavenumber resolution data hence the agent differentiation and false alarm performances of the 16 wavenumber resolution mode can be poorer than that of the 4 wavenumber resolution mode. Therefore a dual "search" and "confirmation" mode approach is used in one embodiment, in which the 16 and the 4 wavenumber resolution modes are used in concert to meet timing and detection requirements. Of course, given faster processors, a single high resolution mode approach will be feasible, or a single mode of suitable resolution may be used. The invention is not limited to 4 or 16 wave number resolution.

Effectively, the search mode operation detects all regions of interest (ROI) that potentially have chemical agents. It must do this with a reasonably low rate of false triggers, but with the same sensitivity as the confirmation mode because to miss a cloud in search-mode is to miss it entirely. A rule is defined such that the search mode can be switched immediately to confirmation mode without scanning the entire field of range. This happens in the mode switch block when the search mode result reaches a high confidence decision that a chemical agent cloud is present. Thus, the processing can detect the chemical agent in the shortest time. The confirmation mode applies a stop and stare operation, in which high resolution (4 cm^"1) data is collected and analyzed to confirm the presence of, and classify the types of chemical agents in the field of view. Any false triggers from the search-mode are rejected. Lower confidence search mode detections are evaluated by the confirmation mode once the field of regard scans have been completed.

A further challenge is that the algorithms must detect down to very low agent signature strengths that approach the noise level of the system with a very low false alarm rate. The small signal detection capabilities are dictated by the concentration and size of the cloud 125, cloud distance and cloud-to-background temperature difference. Furthermore, the small chemical agent signal must be detected under many variations, which could be due to system-to-system difference or changes in operational environment. For example, the frequency of a laser diode that provides the data sampling reference in the sensor varies slightly from one laser to the next. As a result, the spectral resolution may vary from system to system. As another example, the detector response is affected by temperature, and consequently the spectral characteristics will be affected. Extracting the consistent chemical agent spectrum amid the noise and signal variations is critical to the success in the chemical agent detection.

The confirm mode utilizes a sequential decision process whereby a final detection decision is based on N-out-of-M detections from a sequence of confirm mode scans in the same field of view. When a sequential decision is invoked, the final decision at any instance of time can be one of three: "agent detected," "no agent detected," or no "final decision yet." A final "agent decision" is made only when strong evidence of chemical agent is cumulated, such as a majority of the single decisions are agent decision. On the other hand, a final decision on "no agent detected" is made based on very weak or no evidence of chemical agent presence. Thus, any spurious, single scan, false detection will be rejected. In such cases, the detection cycle returns back to the previous stage. No final decision is made when the number of cumulative detected agents does not support nor deny the presence of a chemical agent. If no final decision is made, additional sequential scans are incorporated until an agent or no agent decision is made. The process rules include an upper bound to the value of 'M' as a time constraint. Hence, sequential decision greatly reduces the false alarm rate and increases the confidence that a chemical agent is present when the final "agent detected" decision is made. Once the sequential decision confirms the presence of a chemical agent, the detection cycle switches into the agent cloud mapping stage 240. The agent cloud mapping process locates the extents of the agent cloud based on a search pattern. The search and confirmation modes both process interferograms to make a decision on the presence and class of the chemical agent, if any. Both modes utilize the same algorithm, as shown in FIG. 3, which includes three processes: preprocessing 310, feature extraction 320, and classification 330. Preprocessing transforms the interferogram to the spectral domain and tunes the output to have a common standard free of any sensor and system variation. Feature extraction computes the discriminatory features that are specific to the agent types, interferents, and backgrounds. Classification determines the classes and types of the chemical agents and rejects the interferents and backgrounds. The input data to the two modes differ in resolution. Accordingly the parameters of the algorithms in the two modes also differ. Details of the preprocessing, feature extraction and classification are described with reference to FIG.s 4-7 below.

Preprocessing, the first stage of the detection algorithm, transforms measured interferograms 410 into spectra as illustrated in FIG. 4. The preprocessing stage compensates for any system-to-system variations and drift in time so that the resulting measurement artifacts can be ignored in subsequent algorithm stages. The artifacts that are specifically compensated are frequency-dependent gain, interferogram centerburst position, and spectral resolution. The compensation factors are derived from factory calibration, and calibration functions that are executed at timed intervals, such as every 10 minutes, while in use. One artifact that is not compensated for in the preprocessing stage is the signal-to-noise ratio (SNR) in the spectrum. SNR is addressed in a subsequent stage.

The preprocessing stage consists of the following functions, as shown in FIG. 4. Apodization of the asymmetric interferogram at 420 multiplies the interferogram with a window function shown in block 420. Apodization removes antialiasing due to asymmetry consistent with the well known Mertz method of processing interferograms.

A chirp- Fast Fourier Transform (FFT) 430 calculates spectra to an identical frequency comb of 4 wavenumbers for all systems. Each sensor may have a different sampling reference. The chirp-FFT 430 allows sampling of data at selected frequencies, and interpolates to a selected frequency comb to calibrate between the sensors.

Frequency dependent gain/offset correction is applied at 440 to provide the spectrum 450 comprising an amplitude for each wavenumber. The gain/offset correction is derived from a calibration process. In FIG. 1, a known IR source 180 is periodically inserted between the window 132 and second lens 136 to block out all ambient IR, and provide a known IR radiation.. Gain and offset are calculated that result in an output spectrum matching the source 180, such as a smooth black body. In one embodiment, two known sources represented by source 180 are utilized to determine the gain and offset correction. In a further embodiment, one source is used, and values for a second source are estimated.

The feature extraction stage is shown in FIG. 5, and transforms each spectrum output 520 by the preprocessor into a vector of salient features 540 for a following classifier stage. The goal of the feature extractor 530 is to 1) reduce the quantity of data that must be passed to a classifier for each scene, and 2) transform the scene spectrum to a representation where classification is simpler.

In one embodiment, the spectrum 520 first undergoes a normalization process, wherein the spectrum is divided by a Planck's function, whose temperature is estimated from several points of the spectrum 520. The output is a normalized spectrum, which has peaks and valleys around nominal values of one.

The feature extractor is designed to be sensitive to peaks and valleys in the spectrum. When the system is aimed at a blackbody scene, all elements in the resulting feature vector are zero except for noise. Warm agent clouds relative to the scene produce emission peaks in the spectrum and corresponding positive amplitude feature vector elements. Cool agent clouds produce absorption valleys and negative amplitude feature elements. Each feature in the feature vector is extracted from the spectrum by comparison to a shape template that has been tailored to a particular peak in the absorption coefficient curve of an agent or common interferent. Thus, each feature in the feature vector characterizes the peak or valley in the scene spectrum at a frequency band with a shape template that corresponds to a known agent/interferent absorption feature.

The shape templates utilized by the feature extractor 530 are selected using a heuristic approach with the objective of maximizing detection sensitivity and discriminating capability. The most prominent shape templates from each agent are chosen since these provide the greatest detection sensitivity relative to the noise in the system. Some prominent interferent shape templates are'also chosen because these can sometimes provide discriminating capability. Typically, a set of 20 to 40 shape templates are selected and packaged into a feature matrix 510 that can be loaded into the system via an interface 515.

Typically, a shape template corresponds to a single peak in the absorption coefficient curve of an agent or interferent, although sets of nearby peaks are also sometimes used to form a single template. Also, the shape templates should properly reflect the peak-broadening and other distortion introduced by the measurement system. The feature extractor 530 produces a discriminating feature vector 540. For the subset of scenes that are relatively simple - an agent or interferent cloud against a relatively benign background, an agent decision can be made based on a threshold on the feature vector. For more complex scenes, with multiple agents and/or interferents and feature-rich backgrounds, a further classifier stage described below is utilized. FIG.s 6 A, 6B, 6C and 6D illustrate several filters, or templates of characteristic spectral bands for known agents based on their absorption coefficient curves. Multiple filters for different, or the same agent are shown in templates 610, 615, and 620. The first template 610 comprises two peaks indicated at 625 and 630. Both the height and shape of the curves is representative of the potential agent. Template 615 comprises a curve 635, and template 620 comprises four curves 640, 641, 642, and 643, both of which are representative of agents both by amplitude and shape. Curve 641 and 642 contain a double peak, a small amplitude peak immediately followed by a larger amplitude peak.

Graph 645 illustrates matching of template 615 to detected spectra. Filters 640, 641, 642, and 643 are shown superimposed on the graph with spectral band from the normalized spectra 650. The comparison of the shape templates to the detected spectrum is performed using a least squares fit algorithm, which has been analytically reduced to an equivalent set of matrix operations. The fit algorithm computes amplitude, slope, offset, and mean square error of fit (mse) between the template and the spectral region. Given a shape template, S, whose first and second moment are zero (i.e. mean(S)=0 and S . L =0), and the corresponding spectral region, Y, (both Y and S are vectors of length n), the amplitude, slope, offset, and mse are computed as follows: amplitude = Y . S' slope = Y' . L offset = mean(Y) mse = square_root( (Σ(P(i)-Y(i))²) / n) where L = (Lo - mean (Lo) ) / norm (Lo - mean (Lo)) ; Lo is a vector equals to 1, 2,

..., n

P = offset * U + slope * L + amplitude * S / ΣS(i) ² ; and U is a vector whose n elements equal to 1.

Note that P is the best fit to Y given the free parameters amplitude, slope and offset.

The third stage of the detection algorithm is the classification algorithm 700, as shown in FIG. 7. The main objective of this stage is to classify the extracted feature vector 540 into one or more classes; each class indicates the presence of the associated class agent or the no-agent class. The classifier's challenge is to detect an agent under emission as well as absorption conditions, and also in the presence of different interferents. In this case, the classifier has to map different signatures or feature vectors in order to classify the interferogram properly.

The feature vector is represented at 705 in FIG. 7. A plurality of classifier predetermined parameters for agents are illustrated at 710, and are used to effectively tailor the algorithm to detect the agents. The parameters are provided to a plurality of algorithms comprising feature indices for each classifier, noise threshold 720, feature normalization 730 and neural net classifier 740. Each of these algorithms is duplicated for each different agent to be detected, as indicated with dots, and blocks 755, 760 and 770. The classifier parameters 710 are used to program each of the sets of algorithms based on extensive training and heuristic data.

The first process of the classifier is a preconditioning step, where the classifier performs a normalization step process 730 ... 760 to be able to detect or classify a wider range of agent signatures. In one embodiment, the normalization step is an option determined by a parameter in the classifier parameters 710. Also involved is a noise threshold test 720 ... 755, which measures and removes very weak signals. The measure is a weighed sum of features that are predefined for each agent classifier, and is compared with a threshold. This threshold is adaptively set according to minimum detection requirements for each agent, the false alarm requirements and the SNR for the system in operation. When the measure does not exceed the threshold, a weak signal and a no-agent detection for that agent classifier is declared without exercising that agent neural network. The heart of the classification algorithm is a neural network bank 740 ... 770, in which each of the neural networks is trained to detect a particular agent and reject other non-similar agents, different interferents and background signatures. The neural network is based on the backpropagation architecture with one hidden layer. The size of the hidden layer was carefully chosen in order to classify the agent under different scenarios and not over generalize the detection scheme. An output threshold 780 is associated with each neural network that is tuned based on detection performance and false alarm rate. Since there are usually multiple templates per agent deriving the key discriminating features for that agent, not all of the features in the feature vector need be run through the neural network for it to arrive at the agent decision. The selected feature indices for each classifier are stored in the classifier parameters 710.

Although the classifier architecture is the same for search and confirmation modes, the vector size, hidden layer size, and trained coefficients can be different as represented by classifier parameters 710. The training for each mode is done based on different objectives, which are related on the agent signature, the detection and false alarm requirements, and noise characteristics of each mode.

The simulation software, which is based on a multi-cloud radiometric model, simulates interferograms for chemical agents, simulants, interferents, and other chemical compunds. The simulated data is used for training and testing. In one embodiment, a large amount of training data is utilized. Random subsets of the training data are used to iteratively train the neural networks 740 ... 770. Sequential training sets are also populated with problematic training information from the set used in the previous iteration.

Conclusion

A standoff chemical agent detector is fully automatic and provides real-time, on the move detection for contamination avoidance and reconnaissance operations on a wide variety of land, air, space and sea platforms. A passive, remote Fourier Transform Infrared (FTIR) specfroscopy is used to detect agent clouds at a distance using only thermal emission from the scene.

Algorithms perform pre-processing, feature extraction and classification steps. The pre-processing step calculates the scene spectrum and corrects for system-to- system variations including gain, offset, spectral artifacts and differences in resolution. A robust feature extraction process computes a set of salient features that are tuned to the characteristic spectral bands of the chemical agents represented by templates, and to those of common interfering gases and particulates. The classification process feeds the feature set into trained neural networks to detect the chemical agents while rej ecting background and interferents.

Claims

1. A method of detecting chemical vapor clouds in a scene, the method comprising: generating an interferogram from received scene spectral information; performing apodization on the interferogram; performing a chirp Fast Fourier Transform or Zoom Fast Fourier Transform on the apodized interferogram; applying a calibration curve to correct gain and/or offset; and matching the gain corrected spectrum to selected chemical shape templates.

2. The method of claim 1 and further comprising: creating feature vector based on the matching to the templates; and processing the feature vector through a neural net to determine whether a chemical vapor cloud is present in the scene.

3. The method of claim 2 wherein the feature vector contains features for different vapor clouds, and wherein the neural net comprises a net for each vapor cloud.

4. The method of claim 3 wherein only selected features for one agent are processed by the corresponding neural net for that agent.

5. The method of claim 1 wherein the matching is performed using a least squares fit algorithm.

6. The method of claim 5 wherein the least squares fit algorithm utilizes amplitude, slope, offset, and mean square error of fit.

7. The method of claim 1 and further comprising search and confirm modes having different spectral resolutions.

8. The method of claim 7 wherein an interferogram is generated in both search and confirm modes, and are processed in a similar manner.

9. A method of detecting chemical vapor clouds (any material that exhibits persistent spectral characteristics) in a scene, the method comprising: generating a preprocessed spectrum from received scene spectral information; extracting features from the preprocessed interferogram corresponding to predefined feature templates representative of chemical vapor clouds; and processing the features to determine whether a chemical vapor cloud is present in the scene.

10. The method of claim 9 wherein the features are processed using a neural net for each agent having a single hidden layer and trained by backpropagation.

11. The method of claim 10 wherein features are extracted based on a least squares fit comparison to the predefined feature templates.

12. The method of claim 11 wherein the least squares fit algorithm utilizes amplitude, slope, offset, and mean square error of fit.

13. The method of claim 9 and further comprising first applying the method to an interferogram having a first, search mode resolution, and then applying the method to an interferogram having a second, confirm mode resolution.

14. The method of claim 13 wherein the search mode comprises a 16 wavenumber resolution, and the confirm mode comprises a 4 wavenumber resolution.

15. The method of claim 13 wherein the confirm mode is applied immediately when the search mode reaches a high confidence detection of the presence of an agent, or following a complete scan of the scene by the search mode where lower confidence search mode detections located.

16. A system for detecting chemical vapor clouds in a scene, the system comprising: means for preprocessing an interferogram from received scene spectral information; means for extracting features from the preprocessed interferogram corresponding to predefined feature templates representative of chemical vapor clouds; and means for processing the features to determine whether a chemical vapor cloud is present in the scene.

17. The system of claim 16 wherein the means for processing comprises a plurality of neural nets, each corresponding to one chemical agent.

18. The system of claim 17 wherein the neural nets are trained by back propagation iteratively employing random training subsets of data from a large training set.

19. The system of claim 18 wherein the random training subsets further include problematic data from a previous random subset.

20. The system of claim 16 wherein the means for generating a preprocessed interferogram comprises: means for apodization of the received interferogram; means for interpolating the interferogram to a selected frequency comb; and means for calibrating the interferogram.

21. The system of claim 20 wherein the means for extracting features comprises means for creating a feature vector based on matching to predetermined chemical vapor cloud templates.

22. The system of claim 21 wherein the feature vector is normalized.

23. The system of claim 20 and further comprising means for determining a least squares fit.

24. A computer readable medium having instructions for causing a computer to perform a method detecting chemical vapor clouds in a scene, the method comprising: generating a preprocessed spectrum from received scene spectral information; extracting features from the preprocessed spectrum corresponding to predefined feature templates representative of chemical vapor clouds; and processing the features to determine whether a chemical vapor cloud is present in the scene.

25. A computer readable medium having instructions for causing a computer to perform a method of detecting chemical vapor clouds in a scene, the method comprising: generating an interferogram from received scene spectral information; performing apodization on the interferogram; performing a chirp Fast Fourier Transform on the apodized interferogram; applying a calibration curve to correct gain and/or offset; and matching the gain/offset corrected spectrum to selected chemical vapor cloud shape templates.

25. The computer readable medium of claim 25, wherein the method further comprises: creating feature vector based on the matching to the templates; and processing the feature vector through a neural net to determine whether a chemical vapor clouds is present in the scene.