Page 39 - AJWEP-v22i3

SpillNet CNN model for oil spill detection

and the use of pressure sensors mounted on pipelines To validate the efficacy of the proposed algorithm,
have been conventionally employed for oil spill its performance is compared against existing state-
detection. However, these methods are often ineffective of-the-art models using a comprehensive set of
and time-consuming. The evolution of remote sensing evaluation metrics. The results demonstrate significant
techniques, particularly synthetic aperture radar (SAR) improvements in detection accuracy, showcasing the
imagery, as an oil spill detection and monitoring tool for potential of this novel CNN approach in enhancing oil
large ocean areas, has proven to be effective due to its spill monitoring efforts.
ability to penetrate cloud cover and operate both day and By advancing the accuracy and efficiency of oil spill
night. SAR uses a microwave sensor that records the detection, this research contributes to the broader goals
1
backscattering signals at different frequencies to form of environmental protection and disaster management.
a two-dimensional image. This method is unaffected by The study has applications in the oil and gas industry,
environmental conditions, such as cloudy weather or where it can assist with environmental protection and
the absence of sunlight, and it is cost-effective, covering oil spill management. It is also useful for making
larger areas. However, there is a risk of misclassification predictions to prevent oil spills and enhancing early
in this approach, which may affect the accuracy of oil detection for more effective responses.
spill detection due to the resemblance of captured oil This paper is structured as follows: Section 1 provides
spill spots to other areas with similar features. a general overview of oil spills and background on
In recent years, advancements in machine learning the state-of-the-art technologies used to mitigate their
(ML), particularly convolutional neural networks impact. Section 2 presents the literature review, covering
(CNNs), have revolutionized image analysis tasks. both traditional and non-traditional approaches to oil
2-4
ML approaches can autonomously extract high-level spill detection, their limitations, and the challenges they
features from SAR images and classify them with face while also identifying the research gaps and the
minimal error. CNNs, which are often used for studying novelty of this study. Section 3 details the methodology
and classifying large datasets, have proven effective in employed, while Section 4 reports and discusses the
oil spill detection. Their high accuracy and precision results obtained. The study concludes with a discussion
make them particularly suitable for classification tasks, of the limitations, recommendations, and directions for
especially when dealing with datasets that contain future research.
similar features. The extraction of high-level image
features by CNNs makes it possible to classify images 2. Literature review
accurately. The ability of CNNs to automatically learn
features from raw data makes them ideal for complex 2.1. Traditional approaches to oil spill detection
image recognition tasks. While several state-of-the-art Historically, oil spill detection has been performed
5
models have been developed for oil spill detection, they using optical and SAR images. While optical sensors
8,9
often face challenges in generalizing to diverse SAR are useful, they are limited by cloud cover and lighting
image conditions and varying spill characteristics. 6,7 conditions. On the other hand, SAR can operate in all
10
This study aims to address these challenges by weather conditions and at any time of day or night,
11
designing and implementing a novel CNN-based making it a preferred choice for monitoring marine
algorithm termed “SpillNet,” specifically tailored for oil environments. Techniques, such as thresholding,
12
spill detection in SAR images. The “SpillNet” model is segmentation, and texture analysis have been employed
9
13
a modified form of the conventional CNN and training to identify oil spills in SAR images. However, these
algorithm that incorporates features such as depthwise methods often require significant human intervention
separable convolutional layers (DSCL), batch and are prone to false positives from look-alikes, such
normalization (BN), and residual connections (RC), as low-wind areas, sea ice, and natural surfactants.
which improve the training process and ensure robust
detection and classification of oil spills. The proposed 2.2. ML and CNNs
algorithm leverages the unique characteristics of SAR ML has introduced automation to oil spill detection,
data and the robust feature extraction capabilities of reducing the reliance on human experts. Early ML
CNNs to enhance detection accuracy. Furthermore, the approaches employed shallow models such as support
study evaluates the model’s performance on both seen vector machines and random forests, which showed
and unseen SAR images to assess its generalization promise but were limited in handling complex image
capabilities. features. The introduction of deep learning, particularly

Volume 22 Issue 3 (2025) 33 doi: 10.36922/ajwep.8282