Page 47 - AJWEP-v22i3

SpillNet CNN model for oil spill detection

the existing models in oil spill classification and employed due to its effective dark spot segmentation,
segmentation tasks. which enhanced the feature extraction process, resulting
Figures 4 and 5 display a comparative analysis of the in a high accuracy of SpillNet for dark spot classification.
accuracy of the different models employed in this study, The model segmented the SAR images into super-
highlighting the superior performance of SpillNet. pixel patches and performed classification based on
Figure 6 shows the predicted areas of oil spills using the extracted features. Areas with deep dark spots were
the SpillNet model. Using the proposed CNN SpillNet identified as oil spill regions, while areas with light dark
model, the dark spots were identified, segmented, and spots were classified as look-alike areas. The developed
classified. This model outperformed other CNN models SpillNet CNN model was able to differentiate between

1
0.9
0.8
Performance value 0.6
0.7
0.5
0.4
0.3
0.2
0.1
0
Linknet FPN PSPNet Unet Spillnet
Accuracy 0.919029 0.90936 0.92469 0.936947 0.946947
Mean IoU 0.46909 0.473193 0.508106 0.481241 0.581241
Error rate 0.081971 0.09064 0.07531 0.063053 0.061053
Models employed
Accuracy Mean IoU Error rate

Figure 4. Comparative analysis of the accuracy, mean Intersection over Union, and error rates of the different
models
Abbreviations: FPN: Feature Pyramid Network; PSPNet: Pyramid Scene Parsing Network.

1
0.9
0.8
0.7
Performance value 0.5
0.6

0.4
0.3
0.2
0.1
0
Linknet FPN PSPNet Unet Spillnet
Mean precision 0.587804 0.552293 0.60834 0.613542 0.623542
Mean recall 0.54806 0.566916 0.587413 0.541559 0.571559
Mean pixel accuracy 0.54606 0.566916 0.587413 0.541559 0.59806
Mean specificity 0.922446 0.923545 0.936614 0.924469 0.944469
Models employed
Mean precision Mean recall Mean pixel accuracy Mean specificity
Figure 5. Comparative analysis of the mean performance metrics of the different models
Abbreviations: FPN: Feature Pyramid Network; PSPNet: Pyramid Scene Parsing Network.

Volume 22 Issue 3 (2025) 41 doi: 10.36922/ajwep.8282