Page 42 - AJWEP-v22i3

Umaha, et al.

regularizer and stabilizer during the training process, preparing the data for CNN. The procedural steps for
allowing for a faster and more stable rate of learning the development and implementation of the proposed
while reducing dependence on weight initialization. In SpillNet model are presented in Figure 1. These steps
addition, the RCs ensure effective training of the deep are divided into three major phases: (i) data collection
networks by providing shortcut paths for gradient flow. (which is further divided into the pre-processing
This prevents gradients from becoming too small as and data storage phases), (ii) model development
they propagate back through the network. The special and implementation (which includes algorithm
feature of the modified CNN model, “SpillNet,” is modification and training, along with the selection of
its ability to analyze SAR images and automatically training resources), and (iii) the output and performance
identify and classify oil spills and look-alikes. It is evaluation phase. In the model development and
sensitive to important surface characteristics of the implementation phase, the images were loaded into the
ground or water, such as texture, color differences, and deep learning environment, and multiple convolution
patterns, allowing it to accurately locate oil spills and layers were applied with filters to extract important
distinguish them from other look-alikes. This reduces features from the images. BN was first applied to ensure
the likelihood of false positives and demonstrates the a good distribution of the dataset, followed by the use of
model’s robust predictive and classification capabilities. the Rectified Linear Unit (ReLU) activation functions
to introduce non-linearity and detect initial features.
3. Methodology Pooling layers were also applied to downsample feature
maps, which were then flattened into a one-dimensional
This section describes the SAR image dataset used, as array. Dense layers were used for classification, with the
well as the data augmentation techniques employed to model classifying the images into oil spill and non-oil
enhance the dataset and the pre-processing steps for spill (including the look-alikes) categories. The final

CNN for oil spill detection

Model development and Output and model
Dataset collection
Implementation evaluation
Data storage Algorithm Training Trained
Preprocessing
for use modification resources SpillNet model
& training
Localisation Integration of TensorFlow
and cropping DSCL, BN, Keras library
and RC
Performance metrics:
Rescaling Google Colab Mean precision,
or resizing Fixing of input Python 3.7 mean recall,
layers, conv1
residual blocks accuracy, error rate,
mean IoU, mean
Radiometric pixel accuracy, and
calibration Downsamples, mean specificity
upsamples,
and conv2
Noise filtering
Training, data
Conversion to augmentation,
real luminosity loss function,
values optimizer, learning
rate, batch size,
and epoch
Figure 1. Procedural steps for the development and implementation of the proposed SpillNet model
Abbreviations: BN: Batch normalization; conv1: Initial convolution layer; conv2: Final convolution layer;
CNN: Convolutional neural network; DSCL: Depthwise separable convolutional layers; IoU: Intersection
over Union; RC: Residual connections.

Volume 22 Issue 3 (2025) 36 doi: 10.36922/ajwep.8282