Page 46 - AJWEP-v22i3

Umaha, et al.

predictions based on the data used, while error rate of 0.944469, with the least error rate of 0.061053. This
represents incorrect predictions or detections. Mean IoU demonstrates the superior performance of the Spillnet
is a performance metric used to evaluate the algorithm’s model. Although the PSPNet achieved the highest
accuracy in making correct detections, while mean mean recall of 0.587413 compared to SpillNet’s mean
pixel accuracy is a segmentation metric that indicates recall of 0.571559, overall, Spillnet outperformed the
the percentage of pixels correctly classified in the other models. For example, the proposed mean IoU
image. This is the ratio of correctly classified pixels to (0.581241) was notably higher than those of FPN
the total number of pixels in the image. Mean specificity (0.473193), PSPNet (0.508106), and UNet (0.481241).
represents the ability of the proposed model to correctly This indicates that the proposed model is better at
predict true negatives for each available category. capturing the boundaries of an oil spill. Furthermore,
The results obtained indicate the performance of the the proposed model showed the lowest error rate of
CNN model in making predictions based on the data 0.061053, reinforcing its robustness in segmentation
used. tasks. These performance metrics for SpillNet were
From Table 3, it can be observed that the proposed superior to those of existing deep-learning CNN
SpillNet model achieved the highest mean precision of segmentation models, as shown in Table 3, including
0.623542, accuracy of 0.946947, mean IoU of 0.58124, UNet, LinkNet, FPN, and PSPNet. These findings
mean pixel accuracy of 0.598060, and mean specificity confirm that the proposed model outperforms

Table 2. Validation results
Model Mean Mean recall Accuracy Error rate Mean IoU Mean pixel Mean
precision accuracy specificity
LinkNet 0.587804 0.548060 0.919029 0.081971 0.469090 0.546060 0.922446
FPN 0.552293 0.566916 0.909360 0.090640 0.473193 0.566916 0.923545
PSPNet 0.608340 0.587413 0.924690 0.075310 0.508106 0.587413 0.936614
UNet 0.613542 0.541559 0.936947 0.063053 0.481241 0.541559 0.924469
SpillNet 0.623542 0.571559 0.946947 0.061053 0.581241 0.598060 0.944469
Abbreviations: FPN: Feature Pyramid Network; IoU: Intersection over Union; PSPNet: Pyramid Scene Parsing Network.

Table 3. Comparative analysis of the developed model and existing convolutional neural network models
used for oil spill detection
References Proposed technique Accuracy/precision (%)
Present study Integration of multiple depthwise separable convolutional layers, batch 94.70
normalization, and residual connections
Das et al. 45 CNN with ReLU and sigmoid activation function, binary cross-entropy, and 99.06
Adam optimizer
Guo et al. 46 CNN with sigmoid activation function 91.33
Hidalgo et al. 47 CNN with ReLU* and softmax activation function, and stochastic gradient 96.64
descent
Cantorna et al. 48 CNN with ReLU and binary cross-entropy 98.30
Zeng and Wang 49 CNN with ReLU and softmax cross-entropy, and Adam optimizer 94.01
Song et al. 50 Integrated CNN and support vector machine with ReLU and softmax 99.19
activation function, and cross-entropy
Kang et al. 51 CNN and transformer models 91.00
Liu et al. 38 Multi-source knowledge graph reasoning approach 27.81
Hamza et al. 52 Conventional CNN with standard convolutions 94.30
Mahmoud et al. 22 Enhanced conventional CNN UNet segmentation model with integrated 94.20
dual attention model
Abbreviations: CNN: Convolutional neural network; ReLu: Rectified Linear Unit.

Volume 22 Issue 3 (2025) 40 doi: 10.36922/ajwep.8282