Page 40 - AJWEP-v22i3

Umaha, et al.

CNNs, has dramatically improved the performance of and synthetic data generation have been proposed as
image classification tasks, including oil spill detection. solutions, these techniques still have limitations. 28
14
CNNs are a class of deep learning models designed Another challenge is the high computational cost
to process grid-like data, such as images. They are associated with training CNNs. Advanced models
29
composed of multiple layers, including convolutional require significant computational resources, which
layers, pooling layers, and fully connected layers. may not be readily available in all research settings.
CNNs automatically learn hierarchical features from In addition, while CNNs excel at detecting oil spills
input images, making them highly effective for complex in homogeneous environments, their performance can
pattern recognition. Typically, ML models identify degrade in heterogeneous or noisy conditions.
and classify dark spots or regions in SAR images
that indicate areas covered by oil on the sea surface, 2.5. Identified research gaps
due to the scattering of the SAR signal from the oil. Existing research has demonstrated some automated
These dark spots or regions could either be oil spills or techniques for oil spill detection. For instance, Daniyan
look-alikes. 13-16 The accuracy of the developed model et al. developed a test rig comprising four flow lines,
30
in differentiating oil spills from look-alikes depends each equipped with pressure sensors and gauges. The
on how well the model is trained for identification and system relies on a microcontroller as the main controller,
classification tasks. The three main steps involved in with pressure transducers for monitoring and detecting
oil spill detection and classification include: (i) training pressure drops below a threshold. However, the potential
the dataset and identifying the dark spots, (ii) extracting of ML approaches and their data computational abilities
31
features from these dark spots, and (iii) training a were not explored. Daniyan et al. also reported the
classifier using the extracted features. development of an inline inspection robot for detecting
Once trained, the model can be used to classify issues such as cracks and corrosion in pipelines to
selected dark spots. prevent oil spills. Despite the reported advances, several
gaps in research remain, particularly in utilizing CNNs
2.3. Applications of CNNs in oil spill detection for oil spill detection. One limitation of the study is the
Several studies have demonstrated the efficacy of CNNs reliance on prototype development without validation
in detecting oil spills from satellite imagery. Fingas and against real datasets and empirical results. In addition,
Brown provided a comprehensive review of oil spill the integration of CNNs with other ML approaches,
15
remote sensing techniques, emphasizing the potential of such as recurrent neural networks for temporal analysis,
CNNs. could enhance detection accuracy over time. Moreover,
More recent work by Mahmoudi Ghara et al. the development of lightweight CNN models capable
8
introduced a novel CNN architecture specifically of running on less powerful hardware would make this
designed for oil spill detection, incorporating attention technology more accessible.
mechanisms to focus on relevant parts of the image. The novelty of this study lies in the development
Their approach significantly reduced false positives, and implementation of a unique CNN-based algorithm
particularly in complex environments with numerous termed “SpillNet,” specifically designed for oil spill
look-alikes. Similarly, Zhan et al. applied transfer detection in SAR images. The proposed algorithm
17
8
learning to leverage pre-trained CNNs, achieving state- leverages the unique characteristics of SAR data and
of-the-art results with limited labeled data. the robust feature extraction capabilities of CNNs to
Table 1 presents some of the ML approaches that improve detection accuracy. The architecture integrates
have been implemented for detecting and classifying oil multiple DSCL, BN, and RC to enhance feature
spills with high accuracy. extraction and learning capabilities. The DSCL enables
the model to learn the relationships between complex
2.4. Challenges and limitations of CNNs in oil spill patterns, such as distinguishing oil spills from their look-
detection alikes, with lower computational complexity compared
Despite the success of CNNs, several challenges remain. to standard convolutions. This reduces time and costs,
One major issue in oil spill detection is the scarcity of making the process more efficient. Moreover, the model
labeled training data, which is essential for training deep breaks down the convolution operations into individual
learning models. Annotating large datasets is expensive channel and spatial (pointwise) information, reducing
and time-consuming. Although data augmentation the parameters needed to train the datasets. This

Volume 22 Issue 3 (2025) 34 doi: 10.36922/ajwep.8282