sensors Article

A New Approach of Oil Spill Detection Using Time-Resolved LIF Combined with Parallel Factors Analysis for Laser Remote Sensing Deqing Liu 1,2 , Xiaoning Luan 1 , Jinjia Guo 1 , Tingwei Cui 2 , Jubai An 3 and Ronger Zheng 1, * 1 2 3

*

Optics and Optoelectronics Laboratory, Ocean University of China, Qingdao 266100, China; [email protected] (D.L.); [email protected] (X.L.); [email protected] (J.G.) First Institute of Oceanography, State Oceanic Administration, Qingdao 266100, China; [email protected] Information Science & Technology College, Dalian Maritime University, Dalian 116026, China; [email protected] Correspondence: [email protected]; Tel.: +86-532-6678-1211

Academic Editor: Radislav A. Potyrailo Received: 3 June 2016; Accepted: 16 August 2016; Published: 23 August 2016

Abstract: In hope of developing a method for oil spill detection in laser remote sensing, a series of refined and crude oil samples were investigated using time-resolved fluorescence in conjunction with parallel factors analysis (PARAFAC). The time resolved emission spectra of those investigated samples were taken by a laser remote sensing system on a laboratory basis with a detection distance of 5 m. Based on the intensity-normalized spectra, both refined and crude oil samples were well classified without overlapping, by the approach of PARAFAC with four parallel factors. Principle component analysis (PCA) has also been operated as a comparison. It turned out that PCA operated well in classification of broad oil type categories, but with severe overlapping among the crude oil samples from different oil wells. Apart from the high correct identification rate, PARAFAC has also real-time capabilities, which is an obvious advantage especially in field applications. The obtained results suggested that the approach of time-resolved fluorescence combined with PARAFAC would be potentially applicable in oil spill field detection and identification. Keywords: oil spill identification; time-resolved fluorescence; PARAFAC; laser remote sensing

1. Introduction With the development of marine petroleum exploitation and transportation, fast growing oil spills have caused serious pollution to the oceanic environment and thus become an imminent problem. In order to deal with it, a rapid and reliable detection of oil spill contaminants is necessary. Fluorescence spectroscopy is one of the most useful methods in oil spill analysis as it allows quick and sensitive acquisition of polycyclic aromatic hydrocarbon (PHA) information in petroleum oil [1,2]. To identify spilled oils and link them to the known oil spill sources is important because it can provide evidence for prosecution and guide oil spill countermeasures [3,4]. Oil identification plays an important role in the oil spill disaster management [5,6]. In the study of oil spill identification, the approaches of fluorescence spectroscopy combined with pattern recognition and statistical methods including principal component analysis (PCA), artificial neural network (ANN), support vector machine (SVM), and parallel factors analysis (PARAFAC), etc. have been frequently reported [2,7,8]. PCA is one of the most widely used multivariate analysis methods [3], which has been used in many classic laser fluorosensor systems, such as Environment Canada’s Scanning Laser Environmental Airborne Fluorosensor (SLEAF), University of Oldenburg’s airborne laser fluorosensor (LFS), and Kuwait Institute for Scientific Research’s laser fluorosensor, Sensors 2016, 16, 1347; doi:10.3390/s16091347

www.mdpi.com/journal/sensors

Sensors 2016, 16, 1347

2 of 12

etc. to process and analyze fluorescence spectra data for the oil classification of the three broad oil type categories: light refined, crude, or heavy refined [9–13]. In addition, the application of multivariate statistical methods, such as PARAFAC, to fluorescence excitation-emission matrixes (EEMs) of petroleum oils had good oil identification performances, though these methods are currently more laboratory based [14–17]. Time-resolved laser-induced fluorescence has also been applied to oil identification [18–23]. In the 1970s, the fluorescence decay studies by Rayner and Szabo have shown that measuring time decay together with the fluorescence spectra could allow for the identification of oil in remote sensing experiments [24], and Hegazi et al. have established a new method for remote fingerprinting of crude oil using time-resolved fluorescence, which can be used to discriminate crude oils of different grades [25]. These studies preliminarily evaluated the feasibility of time-resolved fluorescence in field detection of oil spills. It is well-known that PARAFAC is a multi-way decomposition method that generalizes PCA to higher order arrays, and the utmost advantage of this method is that it can recover pure spectra from multi-way spectra data and estimate components successively [26,27]. However, the application of PARAFAC for time-resolved fluorescence is hardly found in published literature, none other than Selli et al. [28] and Saito et al. [29] who had employed PARAFAC in analyzing the time-resolved fluorescence spectra of the mixtures of PAHs and the metal speciation, respectively. In this paper, the application of time-resolved fluorescence technique in laser remote detection of oil spills has been demonstrated, and a new oil spill detection approach by the combination of time-resolved fluorescence and PARAFAC has also been established. The new method is expected to further improve the feasibility and accuracy of oil spill detection. 2. Apparatus and Methods 2.1. Time-Resolved Fluorescence Apparatus Figure 1 is the schematic diagram of the time-resolved fluorescence experimental setup specially developed for oil spill investigation. The experimental setup was established as a laser remote sensing system to meet the demand of shipborne field detection in the future, and had achieved about 5 m detection in the laboratory. As shown in Figure 1, the axis of the transmitter and receiver are coincident. A micro-pulse nitrogen laser (MNL100, LTB, Berlin, Germany) working at a wavelength of 337.1 nm is used as the excitation source. The laser pulse width is about 3 ns, and the laser pulse energy used in the investigation is 90 µJ, with a repetition of 10 Hz. To increase the interaction area between laser and oil and achieve remote detection, the laser beam is expanded and collimated before reaching the oil slick using a fused silica plano-concave lens (focal length: −50.8 mm) and plano-convex lens (focal length: 100 mm). The spot size of laser after beam expanding is about 10 mm. A transparent glass container was used as the sample cell. The seawater was put into the container, and the prepared oil samples were float on its surface. The oil film induced by the expanded laser beam emit fluorescence, and the returned optical radiation is collected by a Newtonian telescope (FirstScope, CELESTRON, Torrance, CA, USA) of which the aperture is 76 mm and focal ratio F/3.95, and then coupled into a 200 µm diameter optical fiber connected to a Czerny-Turner spectrometer (SR-303i, Andor, Oxford, UK). The spectrometer is equipped with a 150 L·mm−1 grating and an intensified charge coupled device (ICCD) camera (iStar DH720-18F-03, Andor, Oxford, UK), giving a broadband coverage from 338 to 736 nm with a spectral resolution of 0.1 nm. The determination of the spectral band relies on existing knowledge that the fluorescence response of crude oil when excited with an ultraviolet laser ranges from 400 to 650 nm, with peak centers in the 480 nm region [30]. In this system, as shown in Figure 1, the delay generator outputs two TTL signals (CH1 and CH2) with fixed delay time ∆t2 to accomplish the time sequence control of the laser and ICCD camera. CH1 and CH2 are the external trigger signal of the laser and ICCD, respectively. The embedded figure is the control sequence of the system, where ∆t1 is the fixed delay between external trigger signal and laser pulse, t1 is the propagation time from laser output to signal return, and t2 , ∆t3 are the gate pulse

Sensors 2016, 16, 1347

Sensors 2016, 16, 1347

3 of 12

3 of 12

Sensors 2016, 16, 1347 3 of 12 external trigger signal and laser pulse, t1 is the propagation time from laser output to signal return,

delayand andt2gate ICCD, respectively. In order to of record complete time-resolved , Δt3 width are theofgate pulse delay and gate width ICCD, respectively. In order tofluorescence record external trigger signal and laser pulse, t 1 is the propagation time from laser output toin signal return, spectra, the gate pulse of the image intensifier in ICCD was sequentially delayed according to the laser complete time-resolved fluorescence spectra, the gate pulse of the image intensifier ICCD was and t 2, Δt3 are the gate pulse delay and gate width of ICCD, respectively. In order to record pulsesequentially by adjusting the gate pulse delay. addition, toadjusting get higher to noise ratio (SNR) and delayed according to the In laser pulse by thesignal gate pulse delay. In addition, to time complete time-resolved spectra, gate pulse of the image intensifier indelay, ICCD get higher signal to noisefluorescence ratio (SNR) and time the resolution, time-resolved spectra were acquired resolution, the time-resolved spectra were acquired with thethe parameters of 110 ns gate 3was ns gate sequentially delayedofaccording the laser by adjusting gate pulse delay. In addition, to with thens parameters 110 ns gatetodelay, 3 nspulse gate step and 10 nsthe gate width. step and 10 gate width. get higher signal to noise ratio (SNR) and time resolution, the time-resolved spectra were acquired with the parameters of 110 ns gate delay, 3 ns gate step and 10 ns gate width.

Figure 1. Schematic diagram thetime-resolved time-resolved fluorescence fluorescence experimental setup for laser remote Figure 1. Schematic diagram ofof the experimental setup for laser remote sensing investigation of oil samples. sensing investigation of oil samples. Figure 1. Schematic diagram of the time-resolved fluorescence experimental setup for laser remote sensing investigation of oil samples. 2.2. Sample Preparation

2.2. Sample Preparation

oil Preparation samples used in the investigation included the refined oils (Gasoline and Diesel) and 2.2. The Sample

The oiloil samples thedifferent investigation the refined Shi138, oils (Gasoline and Diesel) and crude crude samplesused frominfour wells included (Bo601, Chengbei305, and Zhengqi3) of Shengli The oil samples used in the investigation included the refined oils (Gasoline and Diesel) and oil samples four different (Bo601, Shi138, and of D Shengli oilfield. oilfield.from For convenience, thewells oil samples areChengbei305, coded as shown in Table 1, Zhengqi3) G as gasoline, as diesel, crude oil samples from four different wells (Bo601, Chengbei305, Shi138, and Zhengqi3) of Shengli For convenience, the oilC4samples as Chengbei305, shown in Table 1, Gand as Zhengqi3, gasoline, respectively. D as diesel,The and C1, and C1, C2, C3, and as crudeare oilscoded of Bo601, Shi138, oilfield. ForasAmerican convenience, the oil samples are coded as shown in Zhengqi3, Tableoil1,samples G as gasoline, D listed as diesel, density and Petroleum Institute (API) gravity of and different are also in C2, C3, and C4 crude oils of Bo601, Chengbei305, Shi138, respectively. The density and C1, C2, C3, and C4 as crude oils of Bo601, Chengbei305, Shi138, and Zhengqi3, respectively. The Table 1, and they reflect the grade of oil [31]. To prepare a floating oil slick for remote sensing and American Petroleum Institute (API) gravity of different oil samples are also listed in Table 1, and density and American Petroleum Institute (API) gravity of different oil samples are also investigation, someofseawater was put into aa transparent with black tape stucklisted to thein they reflect the grade oil [31]. To prepare floating oil glass slickacontainer for remote sensing investigation, some Table 1, and they reflect the grade of oil [31]. To prepare floating oil slick for remote sensing bottom to eliminate laser reflection. Then, the oil samples were released into the container to form a seawater was put into a seawater transparent container with glass blackcontainer tape stuck toblack the bottom to eliminate investigation, was glass putseawater, into a transparent tape stuck to oil the floating thin oilsome film on the surface of as shown in Figure 2. For with the heavy weight crude laser reflection. Then, the oil reflection. samples were released into the container tointo form acontainer floatingto thin oilafilm bottom to eliminate laser Then, the oil samples were released the form samples of C3 and C4, additional melting procedure was needed because of their high viscosity. on theThey surface of seawater, shown inofFigure 2. For the heavy weight crude oil samples C3 and floating thin oil film onas the surface seawater, as shown in seawater. Figure 2. For the heavy weightofcrude oil C4, were melted in hot water (~60 °C) before releasing into samples of C3 procedure and C4, additional melting procedure was high needed because They of their highmelted viscosity. additional melting was needed because of their viscosity. were in hot ◦ C) before meltedreleasing in hot water (~60 °C) before releasing into seawater. water They (~60 were into seawater.

Figure 2. The picture of the prepared crude oil sample floating on the surface of sea water. Figure 2. The picture of the prepared crudeoil oilsample sample floating of of seasea water. Figure 2. The picture of the prepared crude floatingon onthe thesurface surface water.

Sensors 2016, 16, 1347

4 of 12

Table 1. Physical and spectral parameters of the oil samples used in the investigation. Density (kg·L−1 )

API Gravity

Emission Peak Location λp (nm)

Refined oil

Gasoline(G) Diesel(D)

0.725 0.835

63.4 38.0

409.3 401.0

Crude oil

Bo601(C1) Chengbei305(C2) Shi138(C3) Zhengqi3(C4)

0.807 0.823 0.905 0.979

43.9 40.3 24.9 13.1

440.5 444.4 445.5 449.4

Oil Samples

2.3. Multi-Way Model: PARAFAC As a multi-way decomposition method, PARAFAC decomposes N-way array into N loading matrices. In this work, the time-resolved fluorescence spectra of different oils were organized in a three-dimensional array X, where the number of oil samples was the first dimension, the number of emission wavelength the second, and the number of delay time the third. During the laser remote sensing investigation, the time-resolved fluorescence spectra were taken at five random detection positions on each of six oil floating samples to form 30 samplings in total. At each detection position, 15 emission spectra were obtained at different delay time t2 from 110 ns to 152 ns with a 3 ns interval. From each emission spectrum, 553 emission intensities given by 553 pixels on the ICCD were selected with the wavelength coverage from 338 to 645 nm. In this way, the array X used in the PARAFAC model was set to be 30 × 553 × 15. The PARAFAC model of X can be written as: F

xijk = ∑ ai f b j f ck f + eijk , f =1

(1)

i = 1, · · · , 30 j = 1, · · · , 553 k = 1, · · · , 15 , where xijk is an element of X, which refers to the intensity of time-resolved fluorescence spectra for the ith sample, at emission wavelength j and delay time k. ai f , b j f , ck f are elements of loading matrices A, B, C, respectively, and also the solutions of PARAFAC model. eijk is the residual. The number of columns F in the loading matrices is the number of factors (components). In our analysis, the loading matrix A corresponds to the score of different oils with different factors, B and C correspond to the fluorescence spectra and temporal profiles, respectively. The flowchart of PARAFAC method used to decompose the time-resolved fluorescence spectra of oil samples is shown in Figure 3. The solution to the PARAFAC model of Equation (1) was achieved by alternating least squares (ALS), and more detailed information can be found in [27]. Time-resolved fluorescence spectra of different oils were organized in a three-dimensional array X. The key point of PARAFAC ALS algorithm is the initialization of loading matrices B and C, which use random starting value, and estimation of the number of factors (components) F. The ALS algorithm will improve the fitness of the PARAFAC model. If the algorithm converges to the minimum of the sum of squared residual (SSR), the least squares solution A, B and C to the model would be obtained. A suitable stopping criterion is important, and a common criterion to use is to stop the iterations when the relative change in fit between two iterations is below a certain value (e.g., 10−6 ) [27]. The convergence criterion used in the work is shown as below: I

SSR =

J

K

2 < 1 × 10−6 . ∑ ∑ ∑ eijk

i =1 j =1 k =1

(2)

Sensors 2016, 16, 1347

5 of 12

Sensors 2016, 16, 1347

5 of 12

Sensors 2016, 16, 1347

5 of 12

Figure 3. The flowchart parallelfactors factors analysis analysis (PARAFAC) used to decompose the the Figure 3. The flowchart of ofparallel (PARAFAC)method method used to decompose time-resolved fluorescence spectra ofoil oilsamples. samples. time-resolved fluorescence spectra of Figure 3. The flowchart of parallel factors analysis (PARAFAC) method used to decompose the time-resolved fluorescence spectra of oil samples.

3. Results and Discussion

3. Results and Discussion

3. Results and Discussion 3.1. Time-Resolved Fluorescence Spectra of Oils

3.1. Time-Resolved Fluorescence Spectra of Oils

Figure 4 is the typical returned 3.1. Time-Resolved Fluorescence Spectra of signal Oils of the laser remote sensing system taken from the

Figure is the typical signal the laser remote sensing system taken fromand theecho gasoline gasoline4 samples. It canreturned be seen that the of returned signal mainly contains oil fluorescence Figure 4 is the typical returned signal of the laser remote sensing system taken from the samples. It can be seen that the returned signal mainly contains oil fluorescence and echo signals signals of laser secondary diffraction, and the water Raman signal can also be detected if the oil of gasoline samples. It can be seen that the returned signal mainly contains oil fluorescence and echo laser slick secondary the time-resolved water Ramanfluorescence signal can also be taken detected thedifferent oil slickoil is thin is thin diffraction, enough. Theand typical spectra fromifsix signals of laser secondary diffraction, and the water Raman signal can also be detected if the oil samples are shown in Figure 5. Itfluorescence is obvious that the fluorescence peak of refinedoil andsamples crude oils are enough. The typical time-resolved spectra taken from six different are shown slick is thin enough. The typical time-resolved fluorescence spectra taken from six different oil different, while the peaks similar betweenpeak differentrefined crude oils, and the oils fluorescence intensity of the in Figure 5. Itare is shown obvious thatare the and crude are different, samples in Figure 5. fluorescence It is obvious that theof fluorescence peak of refined and crude oilswhile are  different crude oils varied with the grade of crude oil. The fluorescence peak location ( ) of each oil peaksdifferent, are similar between different crude oils, different and the crude fluorescence of different crude p intensity while the peaks are similar between oils, andintensity the fluorescence of oils varied with the grade of crude oil. The fluorescence peak location (λ ) of each oil sample was listed in sample was listed in Table 1, and the peak fluorescent intensity of oil samples taken at five random p peak location ( p ) of each oil different crude oils varied with the grade of crude oil. The fluorescence detection positions are shown in Table 2. It reflects that the inhomogeneous distribution of oil slick Tablesample 1, and was the peak intensity offluorescent oil samples taken at randomtaken detection listed fluorescent in Table 1, and the peak intensity of five oil samples at five positions random are onin the water has influences on oilinhomogeneous fluorescence signals. shown Table 2. It reflects that the distribution of oil slick on the water has detection positions are shown in Table 2. It reflects that the inhomogeneous distribution of oilinfluences slick on oilonfluorescence the water hassignals. influences on oil fluorescence signals.

Figure 4. Typical returned signal of gasoline (G) taken by the laser remote sensing system. Figure 4. Typical returned signalofofgasoline gasoline (G) remote sensing system. Figure 4. Typical returned signal (G) taken takenby bythe thelaser laser remote sensing system.

Sensors 2016, 16, 1347 Sensors 2016, 16, 1347

6 of 12 6 of 12

Figure 5. taken from gasoline (G), diesel (D)(D) andand 4 crude oil 5. Typical Typicaltime-resolved time-resolvedfluorescence fluorescencespectra spectra taken from gasoline (G), diesel 4 crude samples (C1–C4), (a,b)(a,b) are the oil spectra of gasoline and diesel respectively, (c–f) are theare crude oil samples (C1–C4), are refined the refined oil spectra of gasoline and diesel respectively, (c–f) the oil spectra of B0601, Chengbei305, Shi138 Shi138 and Zhengqi3 respectively. crude oil spectra of B0601, Chengbei305, and Zhengqi3 respectively.

Table 2. The peak fluorescent fluorescent intensity intensity of of six six oil oil samples samples taken taken at at five five random random detection detectionpositions. positions. Detection Detection Positions Positions Position1 Position1 Position2 Position2 Position3 Position3 Position4 Position4 Position5 Position5 Average Averagestandard Relative deviation (RSD) Relative standard deviation (RSD)

Peak Fluorescent Intensity (104 Counts) Peak Fluorescent Intensity (104 Counts) Gasoline(G) Gasoline(G) Diesel(D) Diesel(D) 8.4 8.4 8.0 8.0 4.6 4.6 5.7 5.7 7.6 7.6 6.9 6.9

23.9%

23.9%

Bo601(C1) Chengbei305(C2) Chengbei305(C2) Shi138(C3) Shi138(C3) Zhengqi3(C4) Zhengqi3(C4) Bo601(C1)

50.8 50.8 48.7 48.7 46.3 46.3 49.6 49.6 48.7 48.7 48.8

53.1 53.1 53.8 53.8 56.1 56.1 48.3 48.3 55.6 55.6 53.4

3.4%

5.8%

3.4%

5.8%

48.8

53.4

23.1 23.1 18.8 18.8 17.6 17.6 17.3 17.3 15.7 15.7 18.5

18.5

15.2%

15.2%

3.9 3.9 4.1 4.1 4.3 4.3 4.5 4.5 4.5 4.5 4.3

4.3

5.6%

5.6%

3.9 3.9 3.8 3.8 3.9 3.9 4.0 4.0 3.7 3.7 3.8

3.8

3.3%

3.3%

3.2. Oil Identification Using PARAFAC 3.2. Oil Identification Using PARAFAC The original time-resolved fluorescence spectra of different oils without any preprocessing were The original time-resolved fluorescence ofThe different oilsofwithout anyispreprocessing decomposed and analyzed using the PARAFACspectra method. first step the method to determine were decomposed and analyzed using the PARAFAC method. The first step of the method to the number of factors (F in Equation (1)). Figure 6 is the result of core consistency obtained inisthe determine the number of factors (F in Equation (1)). Figure 6 is the result of core consistency PARAFAC decomposition of original time-resolved spectral data of different oils. From Figure 6, obtained in thethat PARAFAC of original time-resolved spectral of different oils. it can be seen for F > 2,decomposition the core consistency is close to 0. According todata the core consistency From Figure 6, it can be seen that for F > 2, the core consistency is close to 0. According to the core consistency and SSR in the PARAFAC model, the proper number of factors can be determined as

Sensors 2016, 16, 1347

7 of 12

Sensors 2016, 16, 1347 Sensors Sensors 2016, 2016, 16, 16, 1347 1347

7 of 12 77 of of 12 12

and SSR the PARAFAC model, the proper number of factors can be determined asand two.SSR The two. in The core consistency is a measure of “appropriateness” of a given PARAFAC model, is core two. consistency is of model, and SSR two. The Theiscore core consistency is aa measure measure of of “appropriateness” “appropriateness” of aa given given PARAFAC PARAFAC model, and SSR is is consistency a measure of “appropriateness” of a given PARAFAC model, and SSR is expected to expected to stop decreasing when getting an appropriate number of factors because additional expected expected to to stop stop decreasing decreasing when when getting getting an an appropriate appropriate number number of of factors factors because because additional additional stop decreasing getting an appropriate number of factors additional factors only explain factors onlywhen explain random noises [29]. The results of because the PARAFAC decomposition of factors factors only only explain explain random random noises noises [29]. [29]. The The results results of of the the PARAFAC PARAFAC decomposition decomposition of of time-resolved fluorescence spectra of all six oil samples with the of number of factors fluorescence F = 2 are shown random noises [29]. The results of the PARAFAC decomposition time-resolved spectra time-resolved time-resolved fluorescence fluorescence spectra spectra of of all all six six oil oil samples samples with with the the number number of of factors factors F F == 22 are are shown shown in Figures 7 and 8.with the number of factors F = 2 are shown in Figures 7 and 8. of allin six oil samples Figures 7 and 8. in Figures 7 and 8.

Figure 6. The result of core consistency in the thePARAFAC PARAFACdecomposition decomposition of original spectral Figure 6. The result of core consistencyobtained obtained in of original spectral Figure Figure 6. 6. The The result result of of core core consistency consistency obtained obtained in in the the PARAFAC PARAFAC decomposition decomposition of of original original spectral spectral data of different oils. data data of different oils.oils. of different data of different oils.

Figure 7. Intensities of the decomposed factors (F = 2) of PARAFAC as a function of emission

Figure 7. of the decomposed = 2) of PARAFAC as aa function of Figure 7. Intensities Intensities ofthe thedecomposed decomposed factors factors (F (F as as function of emission emission Figure 7. Intensities factors (F==2)2)ofofPARAFAC PARAFAC a function of emission wavelength (a) andof delay time (b), respectively. wavelength wavelength (a) (a) and and delay delay time time (b), (b), respectively. respectively. wavelength (a) and delay time (b), respectively.

Figure 8. PARAFAC score scatter plot of six oil samples on two factors. Figure 8. 8. PARAFAC ofsix sixoil oilsamples samples two factors. Figure score scatter plot on two factors. Figure 8. PARAFAC PARAFACscore scorescatter scatter plot plot of of six oil samples onon two factors.

Figure 7a,b show the intensity of the decomposed factors (F = 2) of PARAFAC as a function of Figure 7a,b show the intensity of the decomposed factors (F = 2) of PARAFAC as a function of

Figure 7a,b show intensityofofthe thedecomposed decomposed factors (F(F = 2) of PARAFAC as a function of Figure show thethe intensity factors = 2) of PARAFAC a function emission7a,b wavelength and delay time, respectively. As can be seen from Figure 7,astwo main of emission emission wavelength wavelength and and delay delay time, time, respectively. respectively. As As can can be be seen seen from from Figure Figure 7, 7, two two main main emission wavelength andhave delaybeen time, respectively. AsPARAFAC can be seendecomposition from Figure 7, main components components (factors) obtained from the oftwo the time-resolved components components (factors) (factors) have have been been obtained obtained from from the the PARAFAC PARAFAC decomposition decomposition of of the the time-resolved time-resolved (factors) have been obtained from the PARAFAC decomposition of the time-resolved fluorescence spectra of the six oil samples. Compared with the spectral shape of original time-resolved fluorescence spectra of different oils, it is clear that the red dash dotted line in Figure 7a mainly corresponds to

Sensors 2016, 16, 1347

8 of 12

Sensors 2016, 16, 1347

8 of 12

fluorescence spectra of the six oil samples. Compared with the spectral shape of original time-resolved fluorescence spectra of different oils, it is clear that the red dash dotted line in Figure 7a mainly corresponds diesel, while theother blackoil solid line in corresponds to theFigure other 8oil diesel, while the black solid linetocorresponds to the samples the experiment. is samples in the score experiment. Figure 8 issix the scatter plot Itofhas all shown six oil samples on and the the PARAFAC scatter plot of all oilPARAFAC samples onscore the two factors. that D, C1, two factors. It has distinguished, shown that D, C1, and C2 clearly distinguished, whereas andC4) twocan heavy C2 can be clearly whereas G can andbe two heavy crude oil samples (C3Gand not. crude oil samples (C3 andshapes C4) can not. In addition, theC3, spectral shapes and intensities of G,from C3, and In addition, the spectral and intensities of G, and C4 are similar to each other the C4 are similar toof each other oils. from the original spectra of different oils. original spectra different As formerly formerly mentioned, mentioned, initial initial attempts attempts to to use use PARAFAC PARAFAC method for original time-resolved fluorescence spectra spectraofofallall samples without any preprocessing had produced unsatisfactory oiloil samples without any preprocessing had produced unsatisfactory results, results, as some oil samples (such as G, C3, and C4) with similar fluorescence spectral shapes and as some oil samples (such as G, C3, and C4) with similar fluorescence spectral shapes and intensities intensities can not be well distinguished. The separation oil factor samples in was the factor space was can not be well distinguished. The separation of oil samples of in the space seemingly related seemingly related to the fluorescence However, the intensitywas of influenced oil fluorescence was to the fluorescence intensity. However, intensity. the intensity of oil fluorescence by several influenced byasseveral factors, such oil as slick laser thickness, pulse energy, oil slick thickness, coefficient of factors, such laser pulse energy, extinction coefficient ofextinction oil, and target distance, oil, target orderoftothe eliminate effects of theintensity, changesmore in fluorescence intensity, etc. and In order todistance, eliminateetc. the In effects changes the in fluorescence detailed analysis can more detailed canpart. be found in the following part. be found in theanalysis following The The process process and and results results of applying applying the the PARAFAC method to normalized time-resolved fluorescence spectra of different oils are discussed in this part. Since Since the the intensity intensity of of oil fluorescence was influenced by many factors, the time-resolved fluorescence spectra of all six time-resolved fluorescence spectra of all six oil samples samples were intensity-normalized, andthen then organized in array the array X its with × 553 15. Also intensity-normalized, and organized in the X with sizeits stillsize 30 ×still 55330 × 15. Also× according according the core consistency andPARAFAC SSR in themodel, PARAFAC model,ofthe number of factors F used in to the core to consistency and SSR in the the number factors F used in the PARAFAC the PARAFAC model was to be 4, and of thedecomposition PARAFAC decomposition of model was determined to determined be 4, and the results ofthe theresults PARAFAC of normalized normalized time-resolved spectra all six oil samples with the numberFof F=4 time-resolved fluorescencefluorescence spectra of all six oilof samples with the number of factors = 4factors are shown are shown 9inand Figures 9 and 9a,b 10. Figure are theofintensity of the decomposed (F = 4) of in Figures 10. Figure are the9a,b intensity the decomposed factors (F =factors 4) of PARAFAC PARAFAC as of a function emission wavelength time, respectively. canfrom be seen from as a function emissionofwavelength and delay and time,delay respectively. As can beAs seen Figure 9, Figure 9, four main components (factors) werefrom obtained from the PARAFAC decomposition of the four main components (factors) were obtained the PARAFAC decomposition of the normalized normalized time-resolved fluorescence spectra of these oil samples. the time-resolved fluorescence spectra of these oil samples. Compared withCompared the spectralwith shape of spectral original shape spectra different canthe also bedashed seen that dashed line in Figure 9a spectraofoforiginal different oils, itof can also be oils, seen itthat pink linethe in pink Figure 9a mainly corresponds mainly to dotted diesel,line the corresponds blue dash dotted linecomponent corresponds to the oil component which to diesel,corresponds the blue dash to the oil which contains the water Raman contains thethe water Ramanline signal, andasthe dashed linemainly as well as the black solid lineand mainly signal, and red dashed as well the red black solid line correspond to the heavy light correspond torespectively. the heavy and light 10 components, respectively. Figure plots 10 is the PARAFAC score scatter components, Figure is the PARAFAC score scatter of all six oil samples on the plots of all sixAs oilcan samples on the fourfrom factors. can bethe appreciated from the results, the oil samples four factors. be appreciated theAs results, oil samples of gasoline, diesel, and crude of diesel,API andgravities crude oils with different API gravities used in the experiment be well oilsgasoline, with different used in the experiment can be well classified. In addition,can seemingly classified. seemingly convincing oil identification results could becomponents. obtained by As anyshown three convincingInoiladdition, identification results could be obtained by any three of the four of the four components. As shown in Table 3, the correct separation rate in the factor space of oil in Table 3, the correct separation rate in the factor space of oil samples used in the experiment proves samples used in the experiment proves that the coupling of intensity-normalized time-resolved that the coupling of intensity-normalized time-resolved fluorescence spectra with the PARAFAC fluorescence spectra with the method would be effective in oil identification. method would be effective in PARAFAC oil identification.

Figure Figure 9. 9. Intensities Intensities of of the the decomposed decomposed factors factors (F (F == 4)4) of of PARAFAC PARAFAC as as aa function function of of emission emission wavelength time (b), (b), respectively. respectively. wavelength (a) (a) and and delay delay time

Sensors 2016, 16, 1347

9 of 12

Sensors 2016, 16, 1347

Sensors 2016, 16, 1347

9 of 12

9 of 12

Figure 10. PARAFAC score scatter plots of of sixsix oiloil samples is the the score scorescatter scatterplot plot on Figure 10. PARAFAC score scatter plots sampleson onfour fourfactors, factors, (a) (a) is on factor2 factor1, and factor2 and factor3, (b) is the score scatter on factor1, factor2 and factor4,(c)(c)isisthe thescore factor1, factor3, (b) is the score scatter plot plot on factor1, factor2 and factor4, score plot onfactor3 factor2,and factor3 and factor4. scatter plotscatter on factor2, factor4.

3.3.

Figure 10. PARAFAC score scatter plots of six oil samples on four factors, (a) is the score scatter plot on factor1, factor2 and factor3, (b) is the score scatter plot on factor1, factor2 and factor4, (c) is the 3.3. Comparative Analysis of PARAFAC and PCA Comparative Analysis PARAFAC score scatter plot onoffactor2, factor3and and PCA factor4.

Generally, laser fluorosensors were used to measure emission spectra of oil with the range Generally, laserAnalysis fluorosensors were used to measure emission spectra of oil with the range gating 3.3. Comparative of PARAFAC gating in oil spill field detections, andand thePCA types of pollution oil can be real-time classified using the in oil PCA spill algorithm field detections, and the types of pollution oilanalysis can be with real-time classified the PCA [9–12]. In order to make a comparative the approach of using PARAFAC Generally, laser fluorosensors were used to measure emission spectra of oil with the range algorithm [9–12]. In order to make a comparative analysis with the approach of PARAFAC coupled coupled with time-resolved fluorescence, PCA was used to process the fluorescence emission gating in oil spill field detections, and the types of pollution oil can be real-time classified using the with time-resolved fluorescence, PCA was used process the fluorescence emission spectra of the six spectra of the six oil samples investigated in thetoexperiment. PCA algorithm [9–12]. In order to make a comparative analysis with the approach of PARAFAC The fluorescencein emission spectra were obtained at two of the five detection positions on each oil samples investigated the experiment. coupled with time-resolved fluorescence, PCA was used to process the fluorescence emission of the six oil samples, and four spectrawere wereobtained selected from the of time-resolved fluorescence spectraon ineach The fluorescence spectra at two the five detection positions spectra of the six oilemission samples investigated in the experiment. each detection position. As shown in Figure 5, the selected delay times were 113 ns, 116 ns, 119 ns, of the sixThe oil fluorescence samples, and four spectra wereobtained selectedatfrom the fluorescence spectra emission spectra were two of thetime-resolved five detection positions on each and 122 ns, respectively. In this way, the size of data array used in PCA was 48 × 553. When PCA is in each detection position. As shown in Figure 5, from the selected delay times were 113 ns, in 116 ns, of the six oil samples, and four spectra were selected the time-resolved fluorescence spectra used to classify the oil types, the fluorescence spectra data has to be normalized for the oil’s each detection position. As shown in Figure 5, the selected delay times were 113 ns, 116 ns, 119 ns, 119 ns, and 122 ns, respectively. In this way, the size of data array used in PCA was 48 × 553. fluorescence emission depending on various factors such as oil thickness and the distance from and 122isns, respectively. In the this oil way, the size of data array used in PCA was 48to × 553. When PCA is When PCA to classify types, the fluorescence data be normalized for the target, etc.used [2]. Therefore, the fluorescence emission spectra spectra data of all sixhas oil samples used in PCA used to classify the oil types, the fluorescence spectra data has to be normalized for distance the oil’s from oil’s fluorescence emission depending on various factors such as oil thickness and the analysis also need intensity-normalized data preprocessing to eliminate the effects of the changes in fluorescence emission depending on various factors such as oil thickness and the distance from target, etc. [2]. Therefore, the processing fluorescence emission data allthree six oil samples used in PCA fluorescence intensity. The results showedspectra that only the of first principal components target, etc. [2]. Therefore, the fluorescence emission spectra data of all six oil samples used in PCA analysis also need intensity-normalized data variance. preprocessing to eliminate thethree effects the changes (PCs) contributed significantly to the data’s The contribution of the PCsofwere 83.2%, in analysis also need intensity-normalized data preprocessing to eliminate the effects of the changes in 7.9%, and 3% respectively, and the total contribution was more than 94%, hence oil separation was fluorescence intensity. The processing thatonly onlythe the first three principal components fluorescence intensity. The processingresults results showed showed that first three principal components based on these three PCs. Figure 11 is the score scatter plot of different oils on the three PCs. It can be (PCs)(PCs) contributed significantly Thecontribution contribution of the three contributed significantlytotothe thedata’s data’s variance. variance. The of the three PCs PCs werewere 83.2%,83.2%, seen that the approach of PCA coupled with fluorescence emission spectra can differentiate refined 7.9%,7.9%, and and 3% 3% respectively, and the wasmore morethan than 94%, hence oil separation respectively, and thetotal totalcontribution contribution was 94%, hence oil separation was was and crude oils, but it is not available to the crude oil samples investigated in the experiment. As on these three PCs. Figure1111isisthe thescore score scatter oilsoils on on the the three PCs.PCs. It canItbecan be basedbased on these three PCs. Figure scatterplot plotofofdifferent different three shown inthe Table 1, the fluorescence peakwith locations ( p ) of these crude oils were quite close. refined approach PCA coupled fluorescence emission spectra cancan differentiate seen seen that that the approach of of PCA coupled with fluorescence emission spectra differentiate refined and crude it isavailable not available to crude the crude oil samples investigated in the experiment. and crude oils, oils, but itbut is not to the oil samples investigated in the experiment. AsAs shown shown in Table 1, the fluorescence peak locations ( p ) of theseoils crude oils quite were quite in Table 1, the fluorescence peak locations (λp ) of these crude were close.close.

Figure 11. PCA score scatter plot of different oils on three principal components.

Figure 11. PCA score scatter plot of different oils on three principal components.

Figure 11. PCA score scatter plot of different oils on three principal components.

Sensors 2016, 16, 1347

10 of 12

A comparative analysis is shown in Table 3, and the separation results of oil samples using PARAFAC and PCA algorithms for fluorescence spectra data of the oils are listed. It can be seen that both PCA and PARAFAC have real-time capabilities. PARAFAC needs determination of the appropriate number of factors before spectra data decomposition, and then the runtime of PARAFAC used to identify oil types is about four seconds. The computation could be faster by means of dedicated hardware resources. By comparison, PCA method could be commonly used to classify broad oil categories, whereas PARAFAC can not only differentiate refined and crude oil but also different crude oils from even the same oilfield with similar fluorescence spectral shapes and intensities. The coupling of intensity-normalized time-resolved fluorescence spectra with the PARAFAC method shows greater advantage in oil identification. Table 3. Comparison of identification results using parallel factors analysis (PARAFAC) and principal component analysis (PCA) algorithms. PARAFAC Oil Samples

Refined oil

Crude oil 1

G D C1 C2 C3 C4

Spectra Number

Number of Overlapping

5 5 5 5 5 5

0 0 1 0 0 0

PCA Runtime

Spectra Number

Number of Overlapping

Runtime

~4 s 1

8 8 8 8 8 8

0 0 6 10 15 21

~6 s

Excluding the time used for testing the appropriate number of components in PARAFAC model.

4. Conclusions In this study, a new approach of oil spill detection for laser remote sensing was established by using time-resolved fluorescence combined with PARAFAC. A series of refined and crude oil samples were investigated using time-resolved fluorescence in conjunction with PARAFAC. The time resolved emission spectra of those investigated samples were taken by a laser remote sensing system at a laboratory basis with a detection distance of 5 m. Based on the intensity-normalized spectra, both refined and crude oil samples even from the same oilfield with similar fluorescence spectral shapes and intensities were well classified without overlapping, by the approach of PARAFAC with four parallel factors. PCA has also been operated as a comparison. It turned out that PCA operated well in classification of broad oil type categories, but with severe overlapping among the crude oil samples from different oil wells. Apart from the high correct identification rate, PARAFAC has also real-time capabilities, which is an obvious advantage especially in field applications. The obtained results suggested that the approach of time-resolved fluorescence combined with PARAFAC would be potentially applicable in oil spill field detection and identification. In future work, the investigation of oil spill field detection using time-resolved fluorescence coupled with PARAFAC method will be carried out. Acknowledgments: This study was financially supported by the National Natural Science Foundation of China (Grant No. 41406111 and 41476159). The authors would like to thank Shengli Oilfield for providing the crude oil samples. Deqing Liu would like to give his thanks to Zhen Shi for his language checking. Author Contributions: Deqing Liu, Xiaoning Luan, and Ronger Zheng proposed the idea. Deqing Liu, Jinjia Guo, Xiaoning Luan, and Jubai An built the experimental system. Deqing Liu and Xiaoning Luan performed the experiments. Deqing Liu and Tingwei Cui processed and analyzed the data. Deqing Liu wrote the paper. Ronger Zheng supervised the research project and revised the draft. Conflicts of Interest: The authors declare no conflict of interest.

Sensors 2016, 16, 1347

11 of 12

References 1. 2. 3. 4. 5. 6.

7. 8.

9. 10.

11.

12.

13. 14.

15.

16.

17.

18. 19. 20. 21.

Frank, U. A review of fluorescence spectroscopic methods for oil spill source identification. Toxicol. Environ. Chem. 1978, 2, 163–185. [CrossRef] Li, J.; Fuller, S.; Cattle, J.; Way, C.P.; Hibbert, D.B. Matching fluorescence spectra of oil spills with spectra from suspect sources. Anal. Chim. Acta 2004, 514, 51–56. [CrossRef] Wang, Z.; Fingas, M.; Page, D.S. Oil spill identification. J. Chromatogr. A 1999, 843, 369–411. [CrossRef] Fingas, M.; Brown, C. Review of oil spill remote sensing. Mar. Pollut. Bull. 2014, 83, 9–23. [CrossRef] [PubMed] Jha, M.N.; Levy, J.; Gao, Y. Advances in remote sensing for oil spill disaster management: State-of-the-art sensors technology for oil spill surveillance. Sensors 2008, 8, 236–255. [CrossRef] Leifer, I.; Lehr, W.J.; Simecek-Beatty, D.; Bradley, E.; Clark, R.; Dennison, P.; Hu, Y.; Matheson, S.; Jones, C.E.; Holt, B.; et al. State of the art satellite and airborne marine oil spill remote sensing: Application to the BP Deepwater Horizon oil spill. Remote Sens. Environ. 2012, 124, 185–209. [CrossRef] Almhdi, K.M.; Valigi, P.; Gulbinas, V.; Westphal, R.; Reuter, R. Classification with artificial neural networks and support vector machines: Application to oil fluorescence spectra. EARSeL eProc. 2007, 6, 115–129. Alostaz, M.D.; Biggar, K.; Donahue, R.; Hall, G. Petroleum contamination characterization and quantification using fluorescence emission-excitation matrices (EEMs) and parallel factor analysis (PARAFAC). J. Environ. Eng. Sci. 2008, 7, 183–197. [CrossRef] Reuter, R.; Wang, H.; Willkomm, R.; Loquay, K.; Hengstermann, T.; Braun, A. A Laser Fluoresensor for Maritime Surveillance: Measurement of Oil Spills. EARSeL Adv. Remote Sens. 1995, 3, 152–169. James, R.T.B.; Dick, R. Design of algorithms for the real-time airborne detection of littoral oil-spills by laser-induced fluorescence. In Proceedings of the 19th Arctic and Marine Oilspill (AMOP) Technical Seminar, Calgary, AL, Canada, 12–14 June 1996; pp. 1599–1608. Brown, C.E.; Marois, R.; Fingas, M.F.; Choquet, M.; Monchalin, J.P.; Mullin, J.V.; Goodman, R.H. Airborne oil spill sensor testing: Progress and recent developments. In Proceedings of the International Oil Spill Conference, Portland, OR, USA, 23–26 May 2001; pp. 917–921. Quinn, M.F.; Al-Otaibi, A.S.; Sethi, P.S.; Al-Bahrani, F.; Alameddine, O. Measurement and analysis procedures for remote identification of oil spills using a laser fluorosensor. Int. J. Remote Sens. 1994, 15, 2637–2658. [CrossRef] Brown, C.E.; Fingas, M.F. Review of the development of laser fluorosensors for oil spill application. Mar. Pollut. Bull. 2003, 47, 477–484. [CrossRef] Christensen, J.H.; Hansen, A.B.; Mortensen, J.; Andersen, O. Characterization and matching of oil samples using fluorescence spectroscopy and parallel factor analysis. Anal. Chem. 2005, 77, 2210–2217. [CrossRef] [PubMed] Mendoza, W.G.; Riemer, D.D.; Zika, R.G. Application of fluorescence and PARAFAC to assess vertical distribution of subsurface hydrocarbons and dispersant during the Deepwater Horizon oil spill. Environ. Sci. 2013, 15, 1017–1030. [CrossRef] [PubMed] Zhou, Z.; Guo, L.; Shiller, A.M.; Lohrenz, S.E.; Asper, V.L.; Osburn, C.L. Characterization of oil components from the Deepwater Horizon oil spill in the Gulf of Mexico using fluorescence EEM and PARAFAC techniques. Mar. Chem. 2013, 148, 10–21. [CrossRef] Zhou, Z.; Liu, Z.; Guo, L. Chemical evolution of Macondo crude oil during laboratory degradation as characterized by fluorescence EEMs and hydrocarbon composition. Mar. Pollut. Bull. 2013, 66, 164–175. [CrossRef] [PubMed] Abu-Zeid, M.E.; Bhatia, K.S.; Marafi, M.A.; Makdisi, Y.Y.; Amer, M.F. Measurement of fluorescence decay of crude oil: A potential technique to identify oil slicks. Environ. Pollut. 1987, 46, 197–207. [CrossRef] Hegazi, E.; Hamdan, A.; Mastromarino, J. New approach for spectral characterization of crude oil using time-resolved fluorescence spectra. Appl. Spectrosc. 2001, 55, 202–207. [CrossRef] Ryder, A.G.; Glynn, T.J.; Feely, M.; Barwise, A.J.G. Characterization of crude oils using fluorescence lifetime data. Spectrochim. Acta Part A 2002, 58, 1025–1037. [CrossRef] Ryder, A.G. Time-resolved fluorescence spectroscopic study of crude petroleum oils: Influence of chemical composition. Appl. Spectrosc. 2004, 58, 613–623. [CrossRef] [PubMed]

Sensors 2016, 16, 1347

22. 23. 24. 25. 26. 27. 28.

29.

30. 31.

12 of 12

Baszanowska, E.; Otremba, Z.; Rohde, P.; Zielinski, ´ O. Adoption of the time resolved fluorescence to oil type identification. J. KONES 2011, 18, 25–29. Liu, D.Q.; Luan, X.N.; Han, X.S.; Guo, J.J.; An, J.B.; Zheng, R.E. Characterization of time-resolved laser-induced fluorescence from crude oil samples. Spectrosc. Spectr. Anal. 2015, 35, 1582–1586. Rayner, D.M.; Szabo, A.G. Time-resolved laser fluorosensors: A laboratory study of their potential in the remote characterization of oil. Appl. Opt. 1978, 17, 1624–1630. [CrossRef] [PubMed] Hegazi, E.; Hamdan, A.; Mastromarino, J. Remote fingerprinting of crude oil using time-resolved fluorescence spectra. Arabian J. Sci. Eng. 2005, 30, 3–12. Harshman, R.A.; Lundy, M.E. PARAFAC: Parallel factor analysis. Comput. Stat. Data Anal. 1994, 18, 39–72. [CrossRef] Bro, R. PARAFAC. Tutorial and applications. Chemom. Intell. Lab. Syst. 1997, 38, 149–171. [CrossRef] Selli, E.; Zaccaria, C.; Sena, F.; Tomasi, G.; Bidoglio, G. Application of multi-way models to the time-resolved fluorescence of polycyclic aromatic hydrocarbons mixtures in water. Water Res. 2004, 38, 2269–2276. [CrossRef] [PubMed] Saito, T.; Sao, H.; Ishida, K.; Aoyagi, N.; Kimura, T.; Nagasaki, S.; Tanaka, S. Application of parallel factor analysis for time-resolved laser fluorescence spectroscopy: Implication for metal speciation study. Environ. Sci. Technol. 2010, 44, 5055–5060. [CrossRef] [PubMed] Fingas, M. Oil Spill Science and Technology; Gulf Professional Publishing, Elsevier: Amsterdam, The Netherlands, 2010. Pantoja, P.A.; Lopez-Gejo, J.; Le Roux, G.A.; Quina, F.H.; Nascimento, C.A. Prediction of crude oil properties and chemical composition by means of steady-state and time-resolved fluorescence. Energy Fuels 2011, 25, 3598–3604. [CrossRef] © 2016 by the authors; licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC-BY) license (http://creativecommons.org/licenses/by/4.0/).