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

Abstract

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.

1. Introduction

According to the International Maritime Organization, different chemical substances, such as petrochemical products and vegetable oil other than crude oil, which varies in physical and chemical properties [1], are considered colorless, hazardous, and noxious substances (HNSs). Petrochemical products, such as benzene and xylene [2] exhibit a broader range of properties (e.g., dissolving, floating, sinking, evaporation, etc.) and different toxicity levels. These chemicals have both acute and long-term ecological effects and cannot be easily recoverable if spilled in the sea [3]. In contrast, nontoxic vegetable oil, such as palm oil can indirectly harm the marine ecosystem [4,5]. Therefore, HNSs spills are considered one of the major causes of marine pollution, damaging aquatic and on-shore human life and likely interfering with other legitimate uses of marine resources [6,7]. Consequently, specific emergency response measures are required if a spill occurs in the sea.

In the past two decades, the number of HNSs spill accidents increased by 3.5 times, for a significant increase in chemical trading via marine transportation. Harmful environmental and economic threats posed by these incidents have led global environmental authorities and scientific research communities to focus on developing specific response solutions to avert and minimize the risk [8]. To deal with spill accidents, accurate, fast, and on-site evaluation of location and features of the spill enable one to take countermeasures to reduce the hazardous effects on marine ecology and estimate the financial cost of the cleaning process [9]. The manual assessment of the HNSs spills is time-consuming and laborious. Therefore, detecting these spills through automatic target detection and classification techniques has become an important subject.

Previous researches have revealed that HNSs spills have their own discrete physicochemical characteristics different from oil spills, which reflect that the approaches for oil spill imaging may not be suitable for HNSs spills. For HNSs detection, several laboratory techniques have been developed which are limited to on-site detection, such as liquid chromatography mass spectrometry [10], UV spectroscopy [11], and electrochemical methods [12]. Synthetic aperture radar (SAR) is an effective but costly imaging technique for the on-site detection of oil spills [13]. In contrast with SAR imaging, optical imaging allows the monitoring of spills at a relatively lower cost, providing more frequent information [14,15,16]. Oil spills exhibit fluorescence features that enable optical sensors to detect them easily. Some potential HNSs, particularly hydrocarbons, possess similar chemistry as oil. This parallel structure and chemistry allow for the use of optical imaging for HNSs spill detection. Spectral reflectance of these chemicals indicates that UV band imaging is appropriate for chemical spill detection. However, unlike crude oil spills, detection of transparent HNSs spills is very difficult due to the lack of color and thin layer floating on the water due to low viscosity. Limited studies indicate that spectral imaging [17], especially in the ultraviolet (UV) band combined with proper data analysis techniques, has an excellent potential for on-site HNSs identification [18,19].

The advent of convolutional neural networks (CNNs) has revolutionized many machine learning areas, achieving success in a wide range of applications, such as object detection, image classification, and segmentation. To address these applications, researchers have proposed several increasingly sophisticated network structures [20,21,22], which recently demonstrated an astounding execution. Deep convolutional neural networks (DCNNs), a derivative of CNN, have been the subject of extensive research in ocean applications [23].

Meanwhile, numerous researches have been carried out using different DCNN architectures to detect oil spills. Most of these studies are based on patch-based oil spill detection using object detection techniques and semantic and instance segmentation of oil spills in aerial and remote sensing data. A systematic summary of several most recent studies since 2017, based on oil spill detection through remote sensing imagery combined with DCNNs, is presented in

Table 1. Summary of reported studies using the applications of DCNN architectures for oil spill detection.

Year	Task	DCNN Architectures	Image Dataset	References
2017	Pixel-based spill classification	CNN with multiple convolutions and pooling layers	Radarsat-2 (SAR images)	[27]
2018	Object (spill) detection	2-stage CNN	SAR images	[28]
	Semantic segmentation	SegNet	Radarsat-2 (SAR images)	[29]
2019	Pixel-based spill classification	1-dimensional CNN	AVIRIS	[30]
	Semantic segmentation	DeepLabv3	Sentinel-1 (SAR images)	[31]
	Object (spill) detection	Multiscale features DCNN	Airborne hyperspectral images	[32]
2020	Pixel-based spill classification	VGG-16	ERS-1,2, COSMO SkyMed, ENVISAT (SAR images)	[33]
	Pixel-based spill classification	CNN + SVM	Radarsat-2 (SAR images)	[34]
	Semantic segmentation	DeepLab + Fully connected conditional random field	QuickBird, Google Earth, and Worldview	[35]
	Instance segmentation	Mask R-CNN	Sentinel-1 (SAR images)	[36]

.

Presumably, a single study has been reported using applications of DCNNs for HNSs spill detection. Huang et al. (2020) used single spectral imagery combined with Faster R-CNN to detect and classify transparent HNSs floating on water [24]. However, the spills, such as highly volatile HNSs chemicals require faster detection. In this regard, this proposal study extends the applications of UV imaging combined with a lightweight YOLOv3 DCNN model to achieve accurate and swift HNSs spill detection. The critical contributions of the work are highlighted as follows:

• YOLOv3 [25] was modified by replacing the DarkNet53 backbone architecture with MobileNetv2 [26]. Depthwise separable convolution layers were implemented in the backbone architecture to obtain a lightweight YOLOv3 model with fewer parameters.
• Loss function was updated by adding the generalized intersection over union (GIoU) for bounding box regression, and k-means clustering was applied to regenerate the appropriate anchor boxes for enhancement in detection accuracy.
• Finally, the lightweight YOLOv3 was trained and tested on the HNSs dataset, and a comparison in the detection based on UV and RGB images was conducted to validate the proposal’s applicability.

The rest of the paper is organized as follows: Methodology is discussed in Section 2, including the HNSs spill dataset and proposed DCNN model for HNSs spill detection. Experimentation is thoroughly described in Section 3. Detection results are comprehensively outlined in Section 4. A comparison of the HNSs detection models is discussed in Section 5. Finally, in Section 6, the proposed work is concluded, and future research directions and discussions are noted.

2. Methodology

2.1. HNSs Image Dataset

Generally, training a DCNN requires a large dataset containing thousands of images to reduce overfitting in the model and enhance the accuracy. However, unlike oil spill datasets, there is no globally available dataset for HNSs spill detection. Therefore, a distinct and comparatively small HNSs spill dataset constructed by the authors of [24] using a spectral and digital imagery system is used in this study. The dataset includes single spectrum UV and RGB images of three colorless HNSs, benzene, xylene, and palm oil.

2.1.1. Image Acquisition

To obtain diversity in HNSs spill image features, imaging experiments were carried out in three locations: A freshwater lake, canal, and artificial plastic-made pool. A multispectral imaging system consisting of an UVTEC-1000 camera (Indigo, Beijing, China), 75 mm optical lens, and a narrow band-pass filter was used to acquire UV images of the spill. The system generated 8-bit grey level images at 365 nm with a resolution of 2016 × 1296. With the exposure time set to 1/50 s, images were captured at an approximately 30 s delay after the sample chemical was released, allowing for the stabilization of the spill on the water surface. For RGB imaging, a digital camera (a6000, Sony, Tokyo, Japan) with a 16–50 mm Sony lens was used to obtain RBG images of the sample spills with the resolution of 3008 × 2000.

Enhancement in the diversity of imaging helps in the detection approach by improving the generalizability of the detection model. For this purpose, spill imaging was carried out from various locations with the varying viewing geometry. By changing the angle between 0–40°, and imaging distance ranging from 1.5–10 m, images of varying shapes and scales were obtained, respectively. Finally, all of the images were down-sampled to the resolution of 350 × 250 to cope with the limited available computing resources. Sample spill images collected from different locations, with varying ambient conditions, including sun reflection and reflection from surrounding objects, are shown in Figure 1.

2.1.2. Data Augmentation

Better performance of deep neural networks depends on the size of the training dataset. Due to the insufficient number of images in the HNSs dataset, the model tends to overfit during training. Random data augmentation techniques were implemented to overcome the overfitting problem, including flipping of a horizontal and vertical axis, rotation, scaling, and affine transformations to increase the dataset volume. After augmentation, all of the images in the dataset were manually annotated using LabelImg [37] to locate the spill in the image. Quantitative and exploratory data analysis of the HNSs dataset is presented in

Table 2. Quantitative and exploratory data analysis of HNSs sample spills image dataset.

Imaging Model	Spilled Chemical	Number of Images Captured at Different Locations			Total Training Images Augmentation (Yes/No)		Total Testing Images
		Freshwater Lake	Canal	Artificial Pool	No	Yes
UV imaging	Benzene	16	29	16	387	958	60
	Xylene	11	28	31
	Palm oil	53	168	35
RGB imaging	Benzene	09	37	–	468	1096	60
	Xylene	14	40	–
	Palm oil	63	305	–

. Compared to benzene and xylene, the number of sample images of palm oil was larger, resulting in better detection accuracy.

2.2. DCNN Model for HNSs Spill Detection

In this study, the YOLOv3 [25], as described thoroughly in Appendix A, is modified to develop a lightweight detection model. Following the physicochemical behavior of sample HNSs chemicals, a faster detection model was adopted by replacing DarkNet53 in the YOLOv3 framework with a lightweight MobileNetv2 [26] as a backbone architecture for feature extraction. With depthwise separable convolutions and inverted residual modules in the network, the MobileNetv2 architecture has fewer parameters, resulting in faster processing speed and requiring low computational power systems. MobileNetv2 utilizes the idea of linear bottleneck residual block to condense the data flowing through the network, which maintains the representation capability of the model. It has two types of residual blocks: Stride 1 consisting of a residual connection and stride 2 for downsizing, as shown in Figure 2. Each block holds three convolutional layers, an expansion layer with 1 × 1 convolution for uncompressing the input tensor size, followed by 3 × 3 depthwise convolution for data filtering, and finally projection layer with 1 × 1 pointwise convolution for compressing the data to bottleneck output. Each layer has ReLU 6 activation function, except for the projection layer to avoid adding non-linearity. Figure 3 represents the object detection algorithm by YOLOv3 with MobileNetv2 backbone architecture.

Based on regression, the lightweight YOLOv3 is a fully convolutional network that is constructed using depthwise separable convolution. After each convolutional layer, a batch normalization layer is added to improve model convergence speed and solve gradient explosion during backpropagation. In the depthwise separable convolutional network, the standard convolution filter is replaced by the depthwise separable filter and pointwise convolution filter, as shown in Figure 4.

The number of parameters of standard convolution filter and depthwise separable convolution filter can be calculated as follows:

Parameters _{(standard conv)} = D × D × N × M

(1)

Parameters _{(depthwise conv)} = D × D × M + N × M

(2)

$$\frac{{\mathrm{Parameters}}_{ \left(\mathrm{depthwise}\mathrm{conv}\right)}}{{\mathrm{Parameters}}_{ \left(\mathrm{standard}\mathrm{conv}\right)}}=\frac{1}{\mathrm{N}}+\frac{1}{\mathrm{D} \times \mathrm{D}}$$

(3)

Equation (3) represents the compression ratio of standard convolution and depthwise separable convolution filter. N and M represent the output and input feature maps, respectively.

2.2.1. Improved Loss Function

In YOLOv3, the evaluation index of object detection depends on IoU, which indicates the amount that the predicted bounding box overlaps with the ground truth bounding box. However, optimizing the loss function based on IoU calculations has significant complications. Suppose the bounding boxes overlap in differential geometry with the same intersection level. In this case, they will result in precisely the same IoU, but the bounding box regression is different, as shown in Figure 5a. In another case, if the predicted and ground truth bounding boxes do not intersect, then the resulting IoU value and gradient will be zero, which cannot optimize the loss function. IoU does not predict the loss due to the distance between predicted and ground truth bounding boxes. GIoU as a new bounding box regression loss function is adopted to address these shortcomings. The difference between IoU and GIoU is defined in Equations (4) and (5), as follows:

$$\mathrm{IoU}=\frac{\left|\right(\mathrm{A} \cap \mathrm{B}\left)\right|}{\left|\right(\mathrm{A} \cup \mathrm{B}\left)\right|}$$

(4)

$$\mathrm{GIoU}=\mathrm{IoU} -\frac{|\mathrm{C} (\mathrm{A} \cap \mathrm{B}\left)\right|}{\left|\mathrm{C}\right|}$$

(5)

where A and B represent the bounding boxes, and C is the smallest circumscribed rectangle enclosing A and B. Depending on the distance between the bounding boxes, the value of IoU and GIoU ranges from [0, 1] to [−1, 1], respectively. In addition, the value of IoU and GIoU is near 1 when the predicted and ground truth bounding boxes overlap. If there is no overlap, the value of IoU is 0, and GIoU gradually approaches −1. Consequently, GIoU is a reasonable distance measurement index that focuses on non-overlapping areas. Figure 5b describes the GIoU variation of C with better and wrong predictions.

Total loss in the lightweight detection model can be calculated by substituting the bounding box regression loss based on GIoU in the YOLOv3 loss equation as in Appendix A.1, Equation (A2), as follows:

$${\mathrm{Loss}}_{\mathrm{BboxRegression} }=1-\mathrm{GIoU}$$

(6)

$$\begin{gathered} {\mathrm{LightweightYOLOv}3}_{\mathrm{Loss}}={ \mathsf{\lambda }}_{\mathrm{coord}} {\displaystyle \sum }_{\mathrm{i}=0}^{{\mathrm{S}}^2} {\displaystyle \sum }_{\mathrm{j}=0}^{\mathrm{B}}{1}_{\mathrm{ij}}^{\mathrm{obj}}\left(1-\mathrm{GIoU}\right) \\ +{\displaystyle \sum }_{\mathrm{i}=0}^{{\mathrm{S}}^2} {\displaystyle \sum }_{\mathrm{j}=0}^{\mathrm{B}}{1}_{\mathrm{ij}}^{\mathrm{obj}}{\left({\mathrm{C}}_{\mathrm{i}}-{\hat{\mathrm{C}}}_{\mathrm{i}}\right)}^2+{ \mathsf{\lambda }}_{\mathrm{noobj}}{\displaystyle \sum }_{\mathrm{i}=0}^{{\mathrm{S}}^2} {\displaystyle \sum }_{\mathrm{j}=0}^{\mathrm{B}}{1}_{\mathrm{ij}}^{\mathrm{noobj}}{\left({\mathrm{C}}_{\mathrm{i}}-{\hat{\mathrm{C}}}_{\mathrm{i}}\right)}^2 \\ +{\displaystyle \sum }_{\mathrm{i}=0}^{{\mathrm{S}}^2}{1}_{\mathrm{i}}^{\mathrm{obj}}{\displaystyle \sum }_{\mathrm{c}\in \mathrm{classes}}^{\mathrm{B}}{\left({\mathrm{p}}_{\mathrm{i}}\left(\mathrm{c}\right) - {\hat{\mathrm{p}}}_{\mathrm{i}}\left(\mathrm{c}\right)\right)}^2 \end{gathered}$$

(7)

2.2.2. Anchor Box Generation

In YOLOv3, the evaluation index YOLOv3 utilizes the concept of anchor boxes while predicting the bounding box. These anchor boxes directly impact the speed and detection accuracy. There are nine anchor boxes in YOLOv3, which are (10, 13), (16, 30), (33, 23), (30, 61), (62, 45), (59, 119), (116, 90), (156, 198), and (373, 326) for the COCO dataset. In this work, nine anchor boxes produced by k-means clustering are (38, 23), (78, 52), (112, 84), (127, 117), (194, 98), (165, 139), (243, 155), (199, 205), and (297, 237).

3. Experimentation

3.1. Model Training

The experiments were performed by the DL open-source library TensorFlow 1.12, OpenCV 4.1.1, and coding was concluded with the high-level language python 3.5 on Ubuntu 18.04 operating system. The computational system included intel core i-7-7700, GPU GTX1080Ti with 12 GB of memory. Multiscale image resolution was performed during the training of the model, which changed the dataset resolution every 10 batches. The learning rate was set to change gradually during training to enhance the convergence of the model. The optimal values of hyperparameters as shown in

Table 3. Training parameters of HNSs spill detection model training.

Model Training Parameters	Parameter Values
Learning rate	1 × e−4 and 1 × e−6
Total training epoch	300 for the baseline model, 450 for lightweight YOLOv3
Batch size	4 and 6
Image size	320 × 320 to 608 × 608
IoU threshold	0.5
Average decay	0.995
Gradient optimizer	Adam

for training were selected through a grid search. Training loss curves of the models are presented in Figure 6.

3.2. Evaluation Protocols

To evaluate the proposed model for HNSs detection, a performance evaluation is conducted [38], and the following techniques were used to assess the model accuracy, which include IoU, precision (P), recall (R), average precision (AP), and mean average precision (mAP). In addition, to measure the efficiency and speed of the network, the detection time and the frame per second (FPS) were also evaluated.

The prime purpose is to calculate the IoU between the predicted bounding boxes and the ground truth. The test result will be true positive (TP) if IoU is higher than the 50% threshold, and false positive (FP) if the IoU value is less than the threshold. TP is the detection of an object correctly with a positive sample, and FP is the detection of an object negatively by accident of a positive sample. False negative (FN) shows that no targeted object in the image carries a positive sample. The precision (P) and recall (R) are calculated as follows:

$$\mathrm{P}=\frac{\mathrm{TP}}{\mathrm{FP}+\mathrm{TP}}$$

(8)

$$\mathrm{R}=\frac{\mathrm{TP}}{\mathrm{FN}+\mathrm{TP}}$$

(9)

$$\mathrm{AP}=\underset{0}{\overset{1}{{\displaystyle \int }}}\mathrm{P}\left(\mathrm{R}\right)\mathrm{dR}$$

(10)

The mAP is an extension of average precision (AP), where the average precision of every class is calculated.

4. Detection Results of the Proposed Model

Multiple distinct experiments were conducted on RGB and UV images for test detection of HNSs to better understand the behavior and evaluate the performance of the YOLOv3 lightweight model.

4.1. Spill Location Detection

The test images were detected, and the IoU values were evaluated for UV and RGB images. The bounding box produced by the YOLOv3 lightweight model is shown in Figure 7. The aqua color is the ground truth of an object, and the red color is the predicted bounding box. The predicted boxes for three samples can be considered correct, as the average IoU values of all the samples is more than 50%. The model showed average IoU values of 86.57% and 80.43% for UV and RGB images. The average IoU values of detection explain that the bounding box regression loss function based on GIoU better fits the proposed YOLOv3 lightweight model. Moreover, the results conclude that the detection rate of palm oil has significantly better IoU values in both UV and RGB images.

4.2. Evaluation Based on Precision and Recall

The precision–recall curve (PRC) is also one of the fundamental indexes to measure the effectiveness and accuracy of the object detection model. The PRC of sample HNSs, including UV and RGB images with the resolution of 608 × 608, is shown in Figure 8. The curves represent linear and better convergence of the proposed model. Of note, PRCs generated by UV images are overall better than RGB images. In UV images, palm oil with the highest AP scores (95.32%) has a perfect classification effect, followed by benzene (85.48%) and xylene (76.34%).

Similarly, AP scores of palm oil, benzene, and xylene in RBG images are 91.49%, 76.24%, and 74.36%, respectively. The results provided evidence that the precision of the proposed model had a transparent edge at a similar recall rate. Moreover, it specifies that the GIoU loss function has enhanced its performance.

4.3. Evaluation Based on Multiscale Resolution

The multiscale resolution technique is used in the YOLOv3 lightweight model. Significant variation in per class AP, mAP, detection time (D-time), and FPS can be observed from lower resolution to higher resolution, as shown in

Table 4. Analysis of proposed YOLOv3 lightweight model based on multiscale resolution.

Image Size	Per Class AP (%) of UV Images			Per Class AP (%) of RGB Images			UV mAP	RGB mAP	Avg D-Time (ms)	FPS
	Benzene	Xylene	Palm Oil	Benzene	Xylene	Palm Oil
320 × 320	54.76	54.39	90.70	52.03	56.43	85.07	75.25	70.02	8.20	120
352 × 352	58.19	58.79	93.07	49.46	78.23	90.93	72.45	70.20	8.56	117
384 × 384	58.31	59.40	93.27	55.55	61.27	92.08	69.15	68.44	8.93	111
416 × 416	67.39	67.96	94.50	64.08	57.85	93.77	69.83	66.97	10.14	98
448 × 448	68.65	75.43	94.79	69.67	57.28	94.27	74.94	69.37	10.96	91
480 × 480	69.52	69.32	94.63	70.33	79.94	94.78	76.62	69.05	11.52	86
512 × 512	74.51	61.53	95.32	70.89	77.92	94.85	77.27	68.16	12.91	77
544 × 544	76.56	71.35	95.17	70.53	80.76	91.67	79.62	69.62	14.07	70
576 × 576	81.96	66.78	94.87	72.29	68.39	91.01	83.05	74.04	15.49	64
608 × 608	85.48	76.34	95.32	76.24	74.36	91.49	86.13	80.60	17.78	57

. This study provides substantial evidence of the influence of resolution on the detection behavior of the model. Moreover, it can be observed that overall detection results are improved using UV imaging.

4.4. Sample HNSs Spill Classification

Figure 9 shows the classification results of three kinds of transparent HNSs spills in UV and RGB images by the YOLOv3 lightweight model. Sample spills in these images can be detected as undistinguishable targets via conventional procedures. The proposed model can classify the transparent spill with advanced imaging techniques and a robust feature extraction algorithm. Although the classification accuracy may differ a little in different circumstances, the sample images carry unalike ambient conditions containing similar objects, such as wave reflection and sun glitter, that are not detected as real targets.

In UV imaging, the transparent spill target has a more significant difference to the background that is ultimately helpful for the detection capabilities of the model. Moreover, the results proved that UV images attained more exceptional accuracy than the detection and classification of target spill in RGB images.

5. Discussion

To better evaluate the proposed YOLOv3 lightweight model, it is compared with the YOLOv3 baseline network trained from scratch with the HNSs dataset using the same training parameters as shown in Table 3. Unlike the YOLOv3 lightweight model, the baseline YOLOv3 network used the IoU regression loss function. A comparative analysis among the three models for HNSs detection is presented in

Table 5. Comparison of proposed YOLOv3 lightweight model.

Characteristic Parameters	Proposed Model	YOLOv3 Baseline	Faster RCNN by the Authors of [24]
mAP (UV)	86.89%	81.13%	86.46%
mAP (RGB)	72.40%	66.94%	66.73%
Parameters (million)	31	61	–
FPS	57	23	5
Average detection time (s)	0.0119	0.0316	0.607
Single Checkpoint size (Megabytes)	107.6	985.1	–

. The results show that the proposed YOLOv3 lightweight model is computationally cost-effective, consuming about 9 times smaller data than the YOLOv3 baseline model. Further investigations revealed that the improved lightweight model resulted in comparable detection accuracy with faster image processing speed and around 3 times lower detection time than the YOLOv3 baseline model. Moreover, the proposed model surpassed the previous study by the authors of [24] in detection accuracy, FPS, and average detection time.

Although the proposed study has made progress in detecting transparent HNSs spills, it also resulted in false prediction as shown in Figure 9c. The encountered problems in the models’ results can be due to the following reasons:

• Overfitting problem caused by the small size of the dataset resulting in the detection model may not generalize well to unseen features in test images.
• Influence of ambient conditions, which may cause errors in detection. This problem can be solved by enhancing the generalization capability of the detection model by adding more training images.

6. Conclusions

The study proposed an improved and lightweight DCNN model to rapidly detect and classify HNSs spills. Due to the unavailability of publicly accessible data, a distinct and generic HNSs spill dataset is constructed, including UV (at 365 nm) and RGB images of benzene, xylene, and palm oil in different aquatic environments. The collected dataset is further augmented to meet the data volume requirement of the DCNN.

A DCNN named the YOLOv3 lightweight model is suggested, which is a modified version of YOLOv3. DarkNet53 backbone architecture is replaced by lightweight MobileNetv2, and bounding box regression loss based on GIoU is introduced in the network. The experiments show that the model is suitable for HNSs spill detection, which resulted in overall IoU of 82.57% and 68.43%, and mAP of 85.89% and 70.40% for UV and RGB images, respectively. The results also revealed that UV imaging is more apposite for the detection purpose of HNSs spills.

Following the physicochemical behavior of HNSs, which are transparent and highly volatile, the proposed model outperformed benchmark DCNN models in accuracy and detection speed. In addition, the model has 31 million parameters that are half of standard YOLOv3 and occupies 107.6 MB on disk, 9 times less than the YOLOv3 baseline model. The processing rate is 57 FPS, which is more than double of YOLOv3. Moreover, the proposed model is 3 times faster in detection, which rapidly detects spills in 11.9 ms on average.

Furthermore, to the best of our knowledge, the suggested model is one of the few reported studies using DCNN for HNSs spill detection. The model accurately and efficiently detects and classifies transparent HNSs spills in coarse conditions, such as wave reflections, water surface illumination, etc. Ultimately, the model can be utilized for swift detection, not only limited to HNSs spill detection on a large-scale, but also any other phenomenon requiring rapid and efficient detection using the lowest possible computational resources. Future studies will include the extension of the HNSs spill dataset by adding a variety of samples collected through monitoring the marine environment, which will enhance the detection efficiency of the model. Moreover, the proposed model will be implemented and tested on large-scale monitoring and mitigation of marine HNSs pollution.