Page 44 - AJWEP-v22i3

Umaha, et al.

ReLU activation and two additional residual blocks Pandasa were used due to their efficiency in handling
with 256 filters. Downsample 2: A 512-filter 3 × large datasets for numerical computations and data
3 convolution with stride 2, followed by BN and manipulation and analysis. Matplotlib was employed for
ReLU activation data visualization, while OpenCV was used for image
(v) Upsample Layers: Upsample 1: Upsampling by a manipulation. TensorBoard was used to monitor training
factor of two, followed by a 256-filter separable metrics in real-time. In addition, the development and
convolution, BN, and ReLU activation. Upsample training processes were carried out on Google Colab, a
2: Upsampling by a factor of two, followed by a free cloud-based environment by Google for executing
128-filter separable convolution, BN, and ReLU Python code.
activation
(vi) Final convolution layer (conv2): A 1×1 convolution 3.4. Computational resources
with the number of filters equal to the number of The model was trained on an NVIDIA Tesla T4 GPU
classes (five), followed by a softmax activation to available on Google Colab, with up to 12 GB of random
produce the final output segmentation map. access memory. The Python environment was set up
The training process and parameter settings involved with Python 3.7, and all required libraries were installed
the following: through pip. Google Drive was used for storing datasets,
(i) Data augmentation: Techniques such as random checkpoints, and logs, ensuring seamless access and
resizing, horizontal/vertical flipping, and cropping management of training data. Figure 2 presents the
were applied to enhance the training dataset. residual block, while Figure 3A and B illustrate the
(ii) Loss function: Categorical cross-entropy was used architecture of the proposed SpillNet model.
as the loss function, as expressed by Equation I:
3.5. Algorithm for the training process
( )
N Input: SAR images I and ground truth annotations

L (y, ˆ = −y ) ∑ i o lg y 1 (I) G, batch size B = 8, epochs E = 100, learning rate α = 1
y
= i 1
× 10 , seed S.
−4
where y represents the ground truth label, ˆ y is the 1. Load data: I and G into memory.

predicted probability, and N is the number of classes. 2. Initialize parameters:
(iii) Optimizer: The Adam optimizer was selected for S←random seed value
its efficiency and adaptability in handling sparse B←8
gradients. E←100
(iv) Learning rate: The learning rate was set to 1 × α ←1×10 −4
10 , with potential adjustments based on model 3. Split data: (I train , G train ), (I , G )←split I and G into
-4
val
val
convergence during training. training and validation sets.
(v) Batch size: A batch size of eight was used to balance 4. For each epoch e from 1 to E:
memory constraints and training efficiency. Shuffle (I train , G train ) using seed S
(vi) Epochs: The model was trained for 100 epochs, Set steps_per_epoch←[len (I train /B)]
with early stopping if the validation accuracy did Set validation_steps←[len (I /B)]
val
not improve for 20 consecutive epochs. For each mini-batch b from 1 to steps_per_epoch:
(vii) Callbacks: Various callbacks were employed, (I , G )←Get a mini-batch of size B from (I train ,
b
b
including EarlyStopping, ReduceLROnPlateau, G train )
ModelCheckpoint, CSVLogger, and TensorBoard, Data augmentation: I ←Apply random
b
to monitor and adjust the training process. resizing, flipping and cropping to I
(viii) Training steps: The number of steps per epoch was Preprocessing: I ←Normalize and resize I to
b
b
calculated based on the size of the training and 224 × 224 pixels
validation datasets, ensuring all data were utilized. Forward pass: ←SpillNet (I )
b
Compute loss: L←Categorical Cross-Entropy
3.3. Implementation tools and frameworks Loss (G ,)
b
The development and training of the proposed model Backpropagation and update weights: Update
were carried out using the TensorFlow framework and SpillNet weights using
Keras library. For data pre-processing, NumPy and Adam optimizer and learning rate α.

Volume 22 Issue 3 (2025) 38 doi: 10.36922/ajwep.8282