Estimation of Hazardous and Noxious Substance (Toluene) Thickness Using Hyperspectral Remote Sensing

According to the Protocol on Preparedness, Response and Co-operation to Pollution Incidents by Hazardous and Noxious Substances (OPRC-HNS Protocol), hazardous and noxious substances (HNS) is referred to as a chemical substance, except for oil, that impairs the ocean by destroying the environment or are harmful to humans and marine life [1] . The increase in maritime transport of HNSs from the larger and faster ships entails the potential risk of marine HNS spill accidents. Highly toxic HNSs, such as benzene, toluene, and xylene, have serious adverse effects on the human body and various organisms living in the marine ecosystems. Most colorless and transparent HNSs have limitations in visual identification and can cause secondary accidents due to explosions and fires, making human access difficult. So, HNS remote sensing based on satellites and aircraft is important for HNS detection at sea because it has the advantage of monitoring a wide area and observing it at a high resolution. To monitor this HNS spill, we designed and performed a ground HNS spill experiment using a hyperspectral sensor to detect HNS areas and estimate the spill volume. The experiment was conducted at the Centre of Documentation, Research and Experimentation on Accidental Water Pollution (CEDRE) marine pool in BREST, France. The HNS image was obtained by pouring 1 L of toluene into an outdoor marine pool and observing it with a hyperspectral sensor capable of measuring the short-wave infrared (SWIR) channel installed at a height of approximately 12 m ( Figure 1 ). At a height of 12 m, the spatial resolution of the SWIR sensor is 10.3 mm and 5.3 mm in the horizontal and vertical directions, respectively. The pure endmember spectra of toluene and seawater were extracted using principal component analysis (PCA) and N-FINDR. This method has the advantage of not requiring input variables other than the number of pure substances. And a Gaussian mixture model (GMM) was applied to the toluene abundance fraction. This model is useful when the data contain multiple Gaussian distributions. The GMM represents a normally distributed subpopulation within the entire population, and it is possible to automatically learn subpopulations without requiring data on subpopulations [2] , [3] . Considering that the histogram of the abundance fraction corresponding to all pixels constituting the HNS image is composed of two peaks containing toluene and seawater rather than a single peak, the GMM is judged to be more suitable than a single model. Empirical linear calibration based on vicarious radiometric calibration was used as an atmospheric correction method for hyperspectral images. It constructs a linear equation using the pixel value of a specific index, whose reflectance is known in advance and is applied to the total image. Rectangular-shaped target plates with white (95%), gray (50%), and black (5%) reflectance were pre-positioned for inclusion in the hyperspectral image. Consequently, a toluene area of approximately 2.4317 mm 2 was detected according to the 36% criteria suitable for HNS detection. The HNS thickness estimation was based on a three-layer two-beam interference theory model [4] . In this model, the HNS thickness was estimated based on the relationship between the reflectance and transmittance at the interface composed of three layers of air, HNS, and seawater. Because toluene has a maximum extinction coefficient of 1.3055 mm at a wavelength of 1678 nm, the closest 1676.5 nm toluene reflectance image was used for thickness estimation. Considering the detection area and ground resolution, the amount of toluene that leaked was estimated to be 0.9336 L. As the amount of toluene used in the actual ground experiment was 1 L, the accuracy of our estimation is approximately 93.36%. Although we precisely designed and constructed the HNS spill experiment and retrieved the actual amount of spill material with reasonable result, there are many possibility to affect to the estimation of HNS thickness in the complicated marine environment. One of the most important causes may be the wind speed controlling evaporation of the HNS and rapid mixing with ambient seawater. The effect of wind field on the thickness estimation should be further investigated based on more intensive experiment in both marine pool circumstances and real sea environment. In this study, the estimated total toluene volume was slightly reduced compared with the amount of toluene spilled in the experiment. The elapsed time from toluene leakage to observation with the hyperspectral sensor was up to 3 min, and there is a possibility that some toluene had evaporated or dissolved. In addition to the effect of wind speed, the evaporation of toluene is affected by other atmospheric and marine environments, such as air temperature, humidity, temperature difference between the air and sea surface temperatures, sea state including various waves. Previous studies on HNS monitoring based on remote sensing are lacking compared with those on oil spills. Considering these limitations, this study evaluated the applicability of HNS detection and thickness estimation based on oil-like reflectance spectral trends at the SWIR wavelengths. The toluene detection results based on hyperspectral remote sensing are expected to be utilized for monitoring marine HNS accidents and evaluating harmful marine ecosystem environments.