With the worldwide demand for chemicals increasing sharply, the amount of chemicals transported by sea has increased by 3.5 times in the past 20 years because of sea transport’s low costs for carrying large quantities over long distances [1
]. The majority of these chemicals are categorized as hazardous and noxious substances (HNSs), which are defined as any substance other than oil that, if introduced into the marine environment, is likely to create human health hazards, harm living resources and other marine life, damage amenities, and/or interfere with other legitimate uses of the sea [2
]. With the increase in the marine transportation of HNSs among major ports, the risk of HNS spillages is further increased [3
]. HNS spills have their own characteristics that are different from oil spills, especially considering the wide variety of products that may be involved [4
]. Approaches for oil detection may not be suitable for HNS spill detection. In order to deal with a spill accident, accurate and rapid knowledge of the spill area, location, and movement enables one to develop effective and efficient countermeasures and therefore significantly reduce the ecological impacts and costs of cleanup operations [5
]. Manual segmentation of spills from a large number of images is time-consuming and laborious. Therefore, automatically detecting those spill regions becomes an important issue.
Unlike other target segmentation tasks, chemical spills have their own challenges. First, in the list of the top 10 chemicals that are likely to pose the highest risk of being involved in an HNS incident [6
], eight chemicals are colorless in the visible spectrum, meaning that the difference in the color between the leaked area and the water surface is small and difficult to identify. Second, most liquid chemicals have low viscosities and low surface tension forces, which lead to a thin liquid film on the water surface [7
] and invalidate the detection method using the changes in surface roughness, such as synthetic aperture radar (SAR) [8
]. Third, due to container ports being frequent accident occurrence places due to their complex environment, accident images usually record the spill target and background, including surface waves, sun reflections, and the shadows caused by buildings, ships, equipment, and clouds. All these challenges make it difficult to recognize chemical leakage targets from the background. To overcome these problems, several relative studies have focused on sensors and automatic detection methods.
With respect to sensors, an HNS monitoring with radar remote sensing experimental study showed that some chemicals are undetectable using SAR images [8
]. This seems to be caused by the high volatility of the tested products and the relatively long time lag between the discharge and observations. Compared to the use of SAR images, optical images allow for a wide swath monitoring and relatively low costs, thereby providing more frequent information [9
]. It is the optical characteristic and chemical properties of oil that makes it possible to detect oil spills using optical sensors [9
]. Some chemical products in HNSs, especially hydrocarbon chemicals, have similar chemistries (OH, CN, and CH bonds) with oil, thereby indicating that there is the potential to use optical images to detect these spills. In the area of oil spill detection, ultraviolet (UV) sensors have been proven to be able to detect thin layers (<0.1 μm) [12
], thereby indicating its potential for detecting thin chemical layers. Recently, unmanned aerial vehicles (UAVs) are becoming popular in coastal environmental monitoring, such as litter, pollution, beach erosion, land use, and anthropological impacts [13
], because of their flexibility and low cost. UAVs equipped with UV sensors have a great potential for the real-time detection of floating HNSs.
Compared to choosing detection hardware, automatic target segmentation is a more challenging task in image understanding and computer vision due to the variety, low contrast, and uneven illumination of chemical spill images. Many methods of image segmentation have been proposed to detect targets on water surfaces, including thresholding [17
], level sets [18
], active contour models [19
], the Mean-Shift [20
], and neural networks [21
]. The level sets method is a branch of active contour model, which represents contour as the zero level set of a higher dimensional function. The use of this method can provide more flexibility in the implementation of active contours. The main drawbacks are high dependence on initializations and it is time-consuming [22
]. The active contour model is based on an energy minimization scheme. The Chan–Vese model is one of the most representative region-based one, which converts image segmentation into a problem of finding contours. With the homogeneity assumption in an image, this method usually fails to segment image with inhomogeneous intensities [23
], while the images of HNS usually present inhomogeneity. The Mean-Shift is a nonparametric iterative algorithm, which has been used in image segmentation. It builds upon the concept of kernel density estimation (KDE), which may misclassify target pixels into its neighborhood background because of low contrast, which is also common in HNS images. The other disadvantage is its time-consuming nature [24
]. Neural network is an efficient supervised learning method. To obtain a good performance neural network, its parameters should be adjusted by a long time back-propagation training process within a large number of training images. It is also highly sensitive to initial training parameter settings [25
]. Dependence on initialization may affect the generalization and efficiency of segmentation method.
Among these algorithms, thresholding is one of the most popular methods that is used for image segmentation because of its effectiveness and simplicity, and it can assist to make an initial estimation for some complex segmentation methods [26
]. Thresholding methods can be classified into three practical categories:
(1) calculating a global threshold for the whole image, such as Otsu [27
] and maximum entropy [28
(2) adapting an adaptive local threshold, such as fuzzy c-means [29
] and adaptive thresholding [30
(3) considering the spatial local information for classifying the pixels, such as two-dimensional (2D) Otsu, 2D maximum entropy [31
], and spatial fuzzy c-means [32
Category 1 only searches for a global threshold from the gray value information and uses it to divide the image into two regions. This method is susceptible to complex backgrounds, low signal-to-noise ratios (SNRs), and uneven illumination, leading to many segmentation errors. Categories 2 and 3 set the threshold according to local explicit and implicit variants, thereby enhancing the recognition accuracy of pixels. However, all these methods fail in the case of images with low contrast between the target and background, which is common in chemical spill images. To fill this gap in chemical spill detection, we conducted a colorless chemical spill experiment to verify the difference and distinguish between a transparent chemical spill and a clean water surface in UV images. These images were captured simulating the moving of UAV. Based on the characteristics of chemical spill images, an automatic chemical spill segmentation method is proposed in this literature. In our method, we first conduct pre-segmentation processing to smooth the noise while maintaining the target’s sharpness. Second, we propose global background suppression (GBS) to reduce the complexity of the background and adaptive target enhancement to improve the saliency of the target. After the above steps, a local fuzzy thresholding method (LFTM), which is robust in resisting artifacts and noise as that proposed in [33
], is adopted to separate the target and the other regions, and the number of regions in the processed image is determined using histogram analysis. The new method can resist complex backgrounds and is suitable for the detection of weak targets. In our experiment of floating xylene on water, the proposed method achieves promising performance on UV image. The next work is to apply our algorithm on UV images that are captured on UAV.
The rest of this paper is arranged as follows. In Section 2
, the overall proposed methodology is outlined. In Section 3
, we conduct an experiment to obtain colorless xylene spill images and implement our method as described. Performance of the algorithm is discussed. Finally, in Section 4
, conclusions on the method and suggestions to future work are mentioned.
In this paper, we present an effective and robust method for automatically extracting xylene spill target from UV images. We proposed an advanced LFTM, which determines cluster numbers automatically and adaptively based on histogram analysis, as the performance of clustering-based segmentation methods highly relies on the selection of the number of clusters. Combining gray value, gradient value, and entropy value, we designed an ATE incorporated by GBS, by which unapparent spill target regions become detectable in UV images with waves, sun reflections, low contrast, and uneven illumination, and hence improve the segmentation precision using LFTM. The whole workflow in this paper should be seen as a basic frame rather than a closed algorithm, and each step can be displaced or replaced by equivalent operations. Due to the use of GBS and ATE, our method loses some local image details which leads to room for improvement in the recall of our method. This is one of our next research focuses.
Parameters (constth, th1, η2, th3, th4, th5, texture_std) and target selection thresholds are experimentally determined and optimized. The proposed method demonstrated its promising detecting capability on UV images with waves, sun reflections, low contrast, and uneven illumination. The overall detecting precision (indicator of F1) of the proposed algorithm is better than the method using original LFTM and other thresholding methods such as Ostu, max entropy, and CV model. Results on current database show a trend of mild increase in computational complexity, along with the increase of accuracy. We also applied our method on UV images with inteferents, e.g., tissue paper and kelp. The result demonstrates that our method is suitable for the segmentation of images with look-alike objects.
To develop the method, increase in the size of the dataset may be helpful on robustness, parameter optimization, and reduction of computational complexity. In the future, more images of large-scale chemical spills captured by moving platforms, e.g., UAVs will be collected to improve our algorithms for in situ remote sensing detection.