Page 45 - AJWEP-v22i3

SpillNet CNN model for oil spill detection

Figure 2. Residual block
Abbreviations: BN: Batch normalization; ReLu: Rectified Linear Unit.

A dataset and prevent variance issues that could hinder
training. RC enhanced the model’s feature extraction
and learning capabilities, ensuring smoother training of
deep networks.

B
4. Results and discussion
The performance of the developed SpillNet model
Figure 3. The architecture of the proposed SpillNet was evaluated on the dataset. In addition, SpillNet’s
model. (A) Network diagram of the proposed SpillNet performance was compared to existing deep learning
model showing the input, hidden, and output layers. CNN segmentation models, including UNet, LinkNet,
(B) Block diagram of the proposed SpillNet model Feature Pyramid Network (FPN), and Pyramid Scene
Abbreviations: BN: Batch normalization; Conv2D: Parsing Network (PSPNet). The evaluation results are
Final convolution layer; ResBlock: Residual Block. presented and thoroughly analyzed in this section.

5. Validation: Evaluate SpillNet on (I , G ) 4.1. Model validation assessment
val
val
Monitor validation accuracy and loss. Adjust Table 2 presents the detection and segmentation
learning rate if performance plateaus. results of various models, including Spillnet, LinkNet,
6. Model checkpoint: Save the best model based on FPN, PSPNet, and Unet. The table compares their
validation accuracy. performance based on metrics such as mean precision,
Repeat steps 2 – 6 until convergence or reaching E mean recall, accuracy, error rate, mean IoU, mean pixel
epochs. accuracy, and mean specificity. The results demonstrate
Output: Trained SpillNet model. that SpillNet significantly outperforms the other models
The primary difference between the SpillNet model across most of these metrics.
and the conventional CNN models lies in the architecture. In this study, seven performance indicators were
Multiple DSCL, BN, and RC were integrated into employed to evaluate the performance of various models
SpillNet to enhance feature extraction, learning, and employed for detecting oil spills. These indicators
classification capabilities, unlike traditional CNN include: mean precision, mean recall, accuracy,
models. DSCL splits the convolutional layers into error rate, mean IoU, mean pixel accuracy, and mean
multiple layers, enabling the model to learn complex specificity. Mean precision is a measure of the accuracy
relationships, such as distinguishing oil spills and look- of positive predictions, while mean recall reflects the
alike features, with reduced computational complexity completeness of positive predictions. High values for
compared to standard convolutions. This modification both precision and recall are generally desirable when
makes the model more time- and cost-efficient. evaluating the performance of ML algorithms, though
Furthermore, the BN in SpillNet was positioned before there may be a trade-off between the two. Accuracy
the ReLU activation function to ensure a well-distributed indicates the performance of the CNN model in making

Volume 22 Issue 3 (2025) 39 doi: 10.36922/ajwep.8282