Page 41 - AJWEP-v22i3

SpillNet CNN model for oil spill detection

Table 1. Existing machine learning models for oil spill detection and their outcomes
References Method Outcome
Zhan et al. 17 Combination of CNNs with the DBSCAN The proposed model achieves an oil spill class
clustering algorithm pixel accuracy of 89.21%
Zhang et al. 18 An improved mask R-CNN model using a The model achieved a high accuracy of 91.5%
combination of convolutional and pooling in detecting oil spills in SAR images
layers to extract features, followed by fully
connected layers for classification
Ma et al. 19 Deep convolutional neural network The improved DeepLabv3+model, using
ResNet-101 as the backbone network and GN
as the normalization layer, achieved superior
performance with high precision (98.92%) and
an inference time of 0.19 s
Yekeen et al. 20 Convolutional neural network (Mask The proposed model achieved an accuracy of
R-CNN) model employing computer 98.3%
vision instance segmentation for oil spill
detection and classification
Feinauer et al. 21 Enhanced CNN model based on The GoogleNet transfer learning model
GoogleNet and VGG16 combined with achieved an improved outcome with 97.5%
transfer learning training accuracy, 95.6% validation accuracy,
and a training loss of 0.0894
Mahmoud et al. 22 A novel deep learning model based on The proposed model achieved an overall
the Dual Attention Model with enhanced accuracy of 94.2%, outperforming the
conventional UNet segmentation network traditional UNet
Ahmed et al. 23 Segmentation network model implemented The proposed model achieved oil spill
in conditional generative adversarial segmentation with an average accuracy of
network 99.04%, an Intersection over the Union index
of 96.59%, and a precision of 85.24%
Huang et al. 24 Faster R-CNN The model achieved precision and recall
metrics of 89.23% and 89.14%, respectively
Basit et al. 25 UNet convolutional neural network model The mean intersection over union value for all
the classes was 75.70%
Abba et al. 26 Three deep-learning algorithms for satellite InceptionV4 outperformed other algorithms
image classification: ResNet50, VGG19, with a classification accuracy of 97% for
and InceptionV4 cloudy, desert, green areas, and water,
followed by VGG19 at 96% accuracy, and
ResNet50 at 93%
Dehghani-Dehcheshmeh Combination of different loss functions The F1-scores for the individual DeepLabV3+,
et al. 27 on CNN networks and hybrid models of FC-DenseNet, and U-Net network models
DeepLabV3+, FC-DenseNet, and U-Net were 75.08%, 73.94%, and 60.85%,
networks based on vision transformers respectively, while the integrated model of
CNN and ViT produced the best F1-score of
78.48%
Abbreviations: CNNs: Convolutional neural networks; DBSCAN: Density-based spatial clustering of applications with noise; GN: Group
normalization; R-CNN: Region-based convolutional neural network; SAR: Synthetic aperture radar; ViT: Vision Transformer.

approach reduces the probability of overfitting, thereby the model’s generalization. Given the complexity of the
improving the model’s robustness and generalization oil dataset and the possibility of look-alikes, it is crucial
ability. Furthermore, BN normalizes the input across to use an algorithm that is less sensitive to the effects
layers to prevent internal variations, which facilitates of initial weights and prevents fluctuations during
convergence, accelerates the learning rate, and improves training. Thus, the introduction of BN acts as both a

Volume 22 Issue 3 (2025) 35 doi: 10.36922/ajwep.8282