On-site visualized classification of transparent hazards and noxious substances on a water surface by multispectral techniques

Combining Deep Learning with Single-Spectrum UV Imaging for Rapid Detection of HNSs Spills

Vital transportation of hazardous and noxious substances (HNSs) by sea occasionally suffers spill incidents causing perilous mutilations to off-shore and on-shore ecology. Consequently, it is essential to monitor the spilled HNSs rapidly and mitigate the damages in time. Focusing on on-site and early processing, this paper explores the potential of deep learning and single-spectrum ultraviolet imaging (UV) for detecting HNSs spills. Images of three floating HNSs, including benzene, xylene, and palm oil, captured in different natural and artificial aquatic sites were collected. The image dataset involved UV (at 365 nm) and RGB images for training and comparative analysis of the detection system. The You Only Look Once (YOLOv3) deep learning model is modified to balance the higher accuracy and swift detection. With the MobileNetv2 backbone architecture and generalized intersection over union (GIoU) loss function, the model achieved mean IoU values of 86.57% for UV and 82.43% for RGB images. The model yielded a mean average precision (mAP) of 86.89% and 72.40% for UV and RGB images, respectively. The average speed of 57 frames per second (fps) and average detection time of 0.0119 s per image validated the swift performance of the proposed model. The modified deep learning model combined with UV imaging is considered computationally cost-effective resulting in precise detection accuracy and significantly faster detection speed.

Hazardous Noxious Substance Detection Based on Ground Experiment and Hyperspectral Remote Sensing

With an increase in the overseas maritime transport of hazardous and noxious substances (HNSs), HNS-related spill accidents are on the rise. Thus, there is a need to completely understand the physical and chemical properties of HNSs. This can be achieved through establishing a library of spectral characteristics with respect to wavelengths from visible and near-infrared (VNIR) bands to shortwave infrared (SWIR) wavelengths. In this study, a ground HNS measurement experiment was conducted for artificially spilled HNS by using two hyperspectral cameras at VNIR and SWIR wavelengths. Representative HNSs such as styrene and toluene were spilled into an outdoor pool and their spectral characteristics were obtained. The relative ratio of HNS to seawater decreased and increased at 550 nm and showed different constant ratios at the SWIR wavelength. Noise removal and dimensional compression procedures were conducted by applying principal component analysis on HNS hyperspectral images. Pure HNS and seawater endmember spectra were extracted using four spectral mixture techniques—N-FINDR, pixel purity index (PPI), independent component analysis (ICA), and vertex component analysis (VCA). The accuracy of detection values of styrene and toluene through the comparison of the abundance fraction were 99.42% and 99.56%, respectively. The results of this study are useful for spectrum-based HNS detection in marine HNS accidents.

Single spectral imagery and faster R-CNN to identify hazardous and noxious substances spills

The automatic identification (location, segmentation, and classification) by UAV- based optical imaging of spills of transparent floating Hazardous and Noxious Substances (HNS) benefits the on-site response to spill incidents, but it is also challenging. With a focus on the on-site optical imaging of HNS, this study explores the potential of single spectral imaging for HNS identification using the Faster R-CNN architecture. Images at 365 nm (narrow UV band), blue channel images (visible broadband of ∼400-600 nm), and RGB images of typical HNS (benzene, xylene, and palm oil) in different scenarios were studied with and without Faster R-CNN. Faster R-CNN was applied to locate and classify the HNS spills. The segmentation using Faster R-CNN-based methods and the original masking methods, including Otsu, Max entropy, and the local fuzzy thresholding method (LFTM), were investigated to explore the optimal wavelength and corresponding image processing method for the optical imaging of HNS. We also compared the classification and segmentation results of this study with our previously published studies on multispectral and whole spectral images. The results demonstrated that single spectral UV imaging at 365 nm combined with Faster R-CNN has great potential for the automatic identification of transparent HNS floating on the surface of the water. RGB images and images using Faster R-CNN in the blue channel are capable of HNS segmentation.