International Journal of Remote Sensing Applications Volume 4 Issue 1, March 2014 doi: 10.14355/ijrsa.2014.0401.01
www.ijrsa.org
Spectral Unmixing Evaluation for Oil Spill Characterization Vassilia Karathanassi Laboratory of Remote Sensing, School of Rural & Surveying Engineering, National Technical University of Athens, 9 Heroon Polytechniou, 15780 Athens, Greece
[email protected] Abstract Hyperspectral remote sensing exploits the optical properties of materials and provides detailed information about them. From a theoretical point of view, in case of oil spills, it cannot only detect and delineate them, but also provide information about the oil type and oil thickness, significantly contributing at the remediation stage of clean-up. In practice, many factors, either associated with the inherent characteristics of oil spills (oil type, quantity, weathering stage, etc.), or with environmental factors (sea bottom cover and depth, waves, etc.) affect the spectral signature of the oil, set constraints on the effectiveness of hyperspectral methods. In this study, the key factors that enable an airborne hyperspectral campaign to implement effective surveys for oil spill detection and characterization are investigated. Additionally, the study focuses on the assessment of environmental and slick parameters for which spectral unmixing-based methods successfully address the problem of oil spill detection and oil type and thickness estimation. For this purpose, study of the spectral behavior of the oil through laboratory measurements and measurements in the complex marine environment was a prerequisite and has initially been carried out. The results showed that almost all the measured spectral signatures as well as their variations can be extracted as endmembers from synthetic images using the unmixing theory. Consequently, laboratory spectral libraries could enable the labeling procedure during the spectral unmixing application on hyperspectral imagery. Unfortunately, oil spectral measurements implemented in marine environment were significantly different because they were affected either by sea bottom contributions (case of light oil and petroleum products) or by sea state conditions which cause high dispersion of oil and spatial variation in oil spill thickness (case of heavy oil and petroleum products). When airborne hyperspectral imagery is processed, it has been found that transparent clouds significantly affect the efficiency of unmixing methods for thin oil spill detection. Their removal, as well as atmospheric correction is strongly recommended. Applying spectral unmixing-based methods on hyperspectral imagery, oil spill detection is effective even in the marginal case of sheens. The results showed that all types and thicknesses of oils can be detected independently of seawater depth through the differences that their spectral signatures present in the wavelengths between 720 μm and
1000 μm. For oil sheens, a single endmember is usually extracted, which leads to relative thickness estimation. For thicker oil spills, many endmembers are extracted each one corresponding to a different thickness and/or emulsion. Further research based on an extended spectral library of measurements in marine environment should be performed in order to enable spectral unmixing-based methods to accurately estimate the oil type, the oil to water ratio of an emulsion as well the oil thickness. Keywords Hyperspectral; Oil Spills; Spectral Unmixing; Endmembers; Labeling
Introduction Airborne hyperspectral imagery is an effective survey tool and a main source for getting near real-time data in case of serious environmental hazards. In the event of an oil spill, the information provided by airborne remote sensing hyperspectral instruments can be used in order to elaborate an effective environmental oil spill protection and response plan, which could help to reduce the environmental impact and the oil spill and cleanup efforts, as well as to protect human lives (Plaza et al, 2005). Hyperspectral imagery can contain over 200 selected wavelengths of reflected and emitted energy presenting a high increase in spectral resolution, compared to multispectral imagery. With this detailed spectral information, one can identify the spectral signature of oil and outline the extent of oil spills. Oil spectral signatures for different oil concentrations can also be used in order to identify the amount of oil contamination in polluted areas, which is necessary to determine proper cleaning processes. Under this concept, several studies have been carried out. In (Salem et al, 2002), spectral similarity measurements, such as the Spectral Angle Mapper (SAM) were used for the detection of oil spills. Pixels were classified as oil spills if they matched the reference oil spill spectral signatures. This work demonstrated that samples of high concentration of
1
www.ijrsa.org
International Journal of Remote Sensing Applications Volume 4 Issue 1, March 2014
hydrocarbon (crude oils) can be discriminated, while light oils such as heating oil cannot (Salem et al, 2001). In (Salem et al, 2004), the Partial Unmixing (PU) technique has been used for the detection of oil spills. PU computes abundances of mixed pixels for selected endmembers which are the targets, and thus there is no need for complete knowledge of all the endmembers present in an image. However, it was concluded that due to the sensor geometry, the complexity of the ocean and atmospheric conditions, as well as the variability of the oil chemical composition, target and background variability is not negligible and can have an impact on the detection result. Another research (Plaza et al, 2005) exploited the simultaneous use of spatial and spectral information by extended mathematical morphology operations which performed well for the extraction of oil spill pure pixels. In (Alam M.S., 2012), five most widely used target detection algorithms have been evaluated for the identification and tracking of surface and subsurface oil spills in ocean environment. Test results using real life oil spill based hyperspectral image datasets showed that the spectral fringe-adjusted joint transform correlation technique and the constrained energy minimization technique yielded better results compared to alternate techniques. Clark et al (2010) used Tetracorder to identify the spectral signatures of oil for different oil thicknesses and oil/water ratios in the Gulf of Mexico on May 17, 2010. AVIRIS airborne hyperspectral data and measured spectra of oil from the Deepwater Horizon have been processed. Spectral analysis in specific wavelengths (near 1.2, 1.7 and 2.3 μm) where oil presents strong absorption features has been performed for estimating oil thickness and volume. Based on the same data Kokaly R. et al, (2010) presented three ratio based indexes for creating colourcomposite images indicative of thick oil/water emulsions. In (Svejkovsky J. et al, 2008), a real time method to estimate the oil slick thickness of crude oils and fuel oils using a UV-visible multispectral sensor is presented. Ratios between various wavelength channels have been computed for both the known clear water area and all other pixels. The thickness determination was based on the deviation of each available ratio from the “clear water ratio” for the same pair of wavelengths. The method was tested on laboratory and real world data. Specific neural network based and fuzzy classification algorithms have been proposed and showed that oil thickness distributions up to 200-300 μm can be mapped with
2
accuracies of 70%, but for oil films thicker than approximately 500 μm, the oil type must be known or estimated for the algorithm to work. Besides, aaccording to Brown C. and Fingas M. (2009), there are no reliable methods either in the laboratory or the field for accurately measuring oil-on-water slick thickness. Dependence of oil reflectance on oil thickness has also been observed through oil spectroscopy and proved using the radiative transfer theory. In (YingCheng, 2008), the spectral response of various bands (550 nm, 645 nm and from 1150 nm to 2500 nm) was found to be correlated with the number of drops, which composed an artificial oil spill. The change in the spectral response from 400nm to 900nm due to oil spill thickness was also confirmed by Croswell et al (2007). In (Lu et al, 2013), a well-controlled laboratory experiment was designed to simulate spectral responses to different oil slick thicknesses. Spectral resampling and normalization methods were used to reduce the differences in spectral reflectance between the experimental background seawater sample and real background seawater. Fitting the analysis with laboratory experimental data results showed a strong linear relationship between normalized oil slick reflectance and normalized oil slick thickness. In (Otremba Z., 2000), the general theory for optical behaviour of hydrocarbon oil films on water has been presented and dependence of reflectance on the thickness and type of oil which covers a water surface has been proved through analysis of the refractive index. In Wettle et al (2009), the reflectance for two crude oil types has been examined as a function of oil thickness and the theoretical minimum limit of detectability of each oil type was calculated for both HYMAP and Quickbird sensors. Reflectance was directly measured and calculated based on the transmittance model. Results showed a decrease in reflectance with increasing oil thickness mainly in the spectral region from 470 to 670 nm. Specific absorbance features were not observed in the visible spectral region, in accordance with literature (Lammoglia et al, 2011). Lammoglia et al, (2011) have focused their research on oil type and age detection. Seventeen petroleum samples were spectrally characterized and the results indicated that it is possible to spectrally discriminate oils according to their density/viscosity, mainly due to the differences between the spectra from the 800 to 1400 nm and 19002300 nm. Reflectance spectra of 1 day oil/water emulsions have been measured and presented an increase in reflectance with increasing their viscosity.
International Journal of Remote Sensing Applications Volume 4 Issue 1, March 2014
All the aforementioned studies, due to the nature of the problem, either use data obtained under laboratory conditions at various scales or, when using remote sensing imagery data, they are case studies. For example, based on the Deepwater Horizon oil spill, an extensive study of thick oil spill spectral behavior in ocean environment has been carried out based on both laboratory and real world data (Clark et al 2010, Kokaly R. et al, 2010, Leifer et al, 2012), but for thin oil spills and oil/water emulsions (
