Page 48 - AJWEP-v22i3

Umaha, et al.

limited. The model performed well on the validation
dataset, indicating that it can generalize effectively
on unseen data and can be used in real-world oil spill
applications. Nevertheless, there are limitations. While
the proposed model’s mean IoU was competitive, it was
slightly lower than that of the FPN model. This finding
suggests that the proposed model may not always
Figure 6. Identified areas of oil spills using the achieve maximum overlap between its predictions and
convolutional neural network SpillNet model the ground truth segments.
The results obtained by Zhan et al., who combined
17
oil spill areas and look-alike regions, achieving a high the CNNs with the density-based spatial clustering of
accuracy of 0.946947. The appearance of oil spills can applications with a noise clustering algorithm, achieved
vary across different locations depending on factors a pixel accuracy of 89.21%, which is higher than the
such as wind speed, water conditions, nature of the one obtained for the SpillNet model proposed in this
spill, and lighting. Furthermore, SAR images can be study (0.598060). However, the SpillNet model showed
affected by noise and other forms of interference, which greater overall accuracy than the model proposed by
12
may increase false positives and affect the accuracy of Zhan et al. The algorithm developed by Vyas et al.
17
classification. Hence, the ability of the proposed SpillNet for oil spill detection first analyzed the SAR images and
model to distinguish between actual areas of oil spills assigned probabilities to the dark spots to identify an
and their look-alikes represents an improvement over oil spill or a resemblance. The results obtained by Bui
existing studies. Mera et al. indicated light dark spots et al., using a combination of deep learning model and
39
36
may result from low wind speeds rather than oil spills. data augmentation technique based on the generative
Najouri et al. stated that there is a higher probability adversarial network model, showed a mean Intersection
37
of obtaining look-alikes when sea wind speeds are too over Union of 72.49%, with a 2.56% increase in the
high or low, and they recommended a wind speed range mean Intersection over Union after applying data
of 2.08 – 8.33 m/s for accurate oil spill detection. Thus, augmentation. These results indicate the feasibility
to improve the accuracy of predictions, Liu et al. of using ML approaches for oil spill detection. The
38
suggested incorporating auxiliary datasets, such as the predictive model developed by Ukpaka et al. for
40
wind speed and direction data, to improve the model’s pipe leakage and insulation failure detection in natural
predictive capabilities. gas transmission pipelines achieved an accuracy of
92.2%.
4.2. Discussion Ukenedo et al. suggested the implementation of a
41
With an accuracy of 0.946947, the evaluation results Health, Safety, and Environmental management system
demonstrated that the proposed model outperforms the policy in the oil and gas industry to effectively mitigate
existing models in the oil spill segmentation task. unsafe conditions such as crude oil spills. Khaira
The effectiveness of the proposed model in et al. recommended the use of Internet of Things-
42
segmenting oil spill regions was evident from its mean based online condition monitoring tools to promote
IoU and Dice coefficient. By striking a satisfactory early fault detection and enhance the reliability of
balance between precision and recall, the proposed model industrial systems, equipment, and infrastructure, such
effectively identified true positives while minimizing as oil pipelines. Similarly, Waghmare et al. suggested
43
false positives. This balance is important for oil spill employing failure mode effects analysis for failure
detection, as both false alarms and missed detections detection and classification to achieve high reliability in
can have significant impacts. With a high accuracy industrial systems and equipment.
of 0.946947, the model proved to be a robust tool for The automation of oil spill detection using artificial
oil spill detection, capable of correctly classifying the intelligence, as demonstrated in this study, is a proactive
majority of pixels in SAR images. The architectural step toward environmental protection. For example,
design of the proposed model, which included separable Promise et al. highlighted that operations in the
44
convolutional layers and RC, enabled efficient training upstream, midstream, and downstream oil sectors in
and faster convergence. The model’s efficiency was Nigeria are a major cause of environmental pollution,
advantageous for oil spill detection applications, which poses significant threats to the environment,
especially in cases where computational resources are aquatic life, plant life, and soil nutrients.

Volume 22 Issue 3 (2025) 42 doi: 10.36922/ajwep.8282