JOIV : Int.
Inform.
Visualization, 9.
- January 2025 216-223
INTERNATIONAL JOURNAL
ON INFORMATICS VISUALIZATION
INTERNATIONAL
JOURNAL ON
INFORMATICS
VISUALIZATION
journal homepage : w.
org/index.
php/joiv Adaptive Inertia Weight Particle Swarm Optimization for Augmentation Selection in Coral Reef Classification with Convolutional Neural Networks Dwi Puji Prabowo a.
Muhammad Syaifur Rohman a,b.
Rama Aria Megantara a,b.
Dewi Pergiwati a,b,*.
Galuh Wilujeng Saraswati a,b.
Ricardus Anggi Pramunendar a,b.
Guruh Fajar Shidik a,b.
Pulung Nurtantio Andono a,b Faculty of Computer Science.
Universitas Dian Nuswantoro.
Semarang.
Central Java.
Indonesia Research Center for Intelligent Distributed Surveillance and Security.
Universitas Dian Nuswantoro.
Semarang.
Indonesia Corresponding author: *dewi.
pergiwati@dsn.
AbstractAiIndonesia possesses the world's largest aquatic resources, with 17,504 islands and 6.
49 million square kilometers of sea.
Located in the coral triangle.
Indonesia is home to diverse marine life, including vital coral reefs.
However, these reefs face threats from climate change, pollution, and human activities, endangering biodiversity and coastal communities.
Therefore, monitoring and preservation are crucial.
This study evaluates various augmentation methods for classifying underwater coral reef images using Convolutional Neural Networks (CNN.
Effective augmentation methods are essential due to the unique characteristics of these images.
The methodology includes testing different augmentation methods, epoch parameters, and CNN parameters on a coral reef image Five optimization algorithms (AIWPSO.
GA.
GWO.
PSO, and FOX) are compared.
The highest accuracy, 95.
64%, is achieved at the 10th epoch.
AIWPSO and GA show the highest average accuracies, 93.
44%, and 93.
50%, respectively, with no significant performance differences among the algorithms.
Statistical analysis using the Wilcoxon test indicates a significant difference between training and validation accuracy .
-value = 0.
These findings underscore the importance of selecting augmentation methods that align with the characteristics of each optimization algorithm to enhance classification performance.
The results provide valuable insights into improving the quality and diversity of input data for classification algorithms in underwater image analysis.
They highlight the necessity of matching augmentation methods to specific optimization algorithms to boost accuracy and effectiveness significantly.
Future research should explore additional augmentation methods and optimization algorithms further to enhance the robustness and accuracy of underwater image classification.
KeywordsAi Coral Reef Classification.
Image Augmentation.
Particle Swarm Optimization.
Adaptive Inertia Weight.
Optimization Manuscript received 13 May 2024.
revised 7 Aug.
accepted 4 Sep.
Date of publication 31 Jan.
International Journal on Informatics Visualization is licensed under a Creative Commons Attribution-Share Alike 4.
0 International License.
Indonesia and several global nations in safeguarding and conserving their marine resources .
Monitoring and preserving marine ecosystems is crucial globally due to the many dangers facing these diverse organisms.
According to the World Bank's 2020 statistics.
Indonesia is a habitat for 166 fish species at risk of extinction.
In addition, the International Union for Conservation of Nature's (IUCN) Red List identifies 308 aquatic biota species that deserve specific protection.
Meanwhile, the Indonesian Institute of Sciences (LIPI) estimates that approximately 1.
7 million hectares of coral reefs in Indonesian waters are at risk of destruction, posing a threat to the survival of 166 fish species and various other forms of marine life in Indonesian waters.
INTRODUCTION
Indonesia has the most considerable aquatic resources globally, boasting 17,504 islands and a sea area of 6.
49 square kilometers .
Consequently, the country is home to many coral reefs .
, .
Indonesia's geographical location also positions it in the epicenter of the world's coral triangle, a region rich in diverse marine life in Indonesian seas.
Several categories of coral reefs serve as habitats for diverse biota, particularly fish species .
Coral reefs provide a valuable supply of protein and nourishment for marine species while providing a haven for a wide range of aquatic biota.
Nevertheless, this presents a formidable obstacle for Aruna .
uses convolutional neural networks (CNN) techniques to classify coral reef images into healthy and stressed categories.
Through hyperparameter fine-tuning and leveraging pre-trained models like ResNet50 and Inception V3, their custom CNN model achieves up to 90% accuracy.
Gapper et al.
utilized LDA on Landsat8 data achieving 1% accuracy in classifying coral reefs in the Icelandic Sea.
Xu et al.
enhanced CNN with VLAD and PCA, improving coral reef classification accuracy.
Gomez-Rios et .
classified coral reef images based on texture using CNN techniques.
Lumini et al.
fine-tuned CNNs on EILAT and RSMAS datasets, reaching 98% average accuracy.
Shaker et al.
employed DenseNet and MobileNet on various coral reef datasets, enhancing accuracy with imageenhancing techniques.
Sharan et al.
utilized CNN for scleractinian coral classification, achieving up to 94.
accuracy with image enhancement.
Borbon et al.
classified coral reefs into three groups with data from multiple databases, improving accuracy using augmentation methods.
Overall, these studies emphasize the importance of modern image enhancement and augmentation methods in improving coral reef species classification accuracy.
Researchers use CNN to enhance the quality of images via augmentation or alternative approaches.
The augmentation approach yields improved performance in recognizing different species or assessing the health status of coral reefs due to the alterations it introduces to the object's features.
However, most individuals only use an unreasonable selection of augmentation methods from various available Hence, this research suggests employing Particle Swarm Optimization (PSO) to choose the augmentation method from available options.
On the other hand, this research also updates the adaptive inertia weight values in the PSO algorithm.
The PSO approach was selected for its computation time efficiency and comparable performance results to other techniques .
This research suggests combining augmentation approaches to attain the highest accuracy in classifying coral reefs' health.
Furthermore, given the diverse range of augmentation methods, this research introduces a novel performance framework to achieve the utmost precision in categorizing coral reef health.
This study also examines the relationship between image enhancement via augmentation and the accuracy of coral reef health ratings.
The study assesses the ideal parameters in the PSO approach, where updating the inertia weights is changed adaptively.
This aims to get the best mix of augmentation and performance.
This article is composed of several sections.
Section 1 emphasizes the introduction and relevant research.
Section 2 presents research suggestions and experimental design.
The empirical findings and analysis are elucidated in Section 3.
Finally, the conclusion section succinctly summarizes the study's fundamental findings.
The detrimental effects of climate change, pollution, escalating temperatures, and human activities have increased illness and mortality rates among coral reefs.
The abrupt fluctuations in ocean temperature resulting from the El Nino phenomena significantly contribute to coral reef issues.
El Nino is a periodic climatic phenomenon that happens every two to seven years and can potentially induce drought The phenomenon arises when the water temperature is above the threshold that coral reefs can endure, leading to the widespread destruction of these ecosystems in numerous places.
Coral reef mortality leads to a decline in coral abundance and results in the loss of coral reefs.
Temperature conditions are directly linked to the alteration of coral reef coloration, which occurs universally throughout all regions of the planet.
In addition, there are several additional contributing reasons to the demise of coral reefs, including seawater contamination resulting from oil pollution, urban effluent, and other factors such as illicit fishing .
These activities result in significant ecological harm to the physical structure of coral reefs, affecting marine ecosystems.
The proximity of coastal communities that relies on the ecological diversity of coral reefs for their livelihoods substantially influences the well-being of coastal areas.
Consequently, the healthy condition of coral reefs has a lot of beneficial effects on the coastal fisheries, development, and tourist industries.
The issue of coral reefs being destroyed and leading to the endangerment of biota is not limited to Indonesia.
it affects all nations worldwide .
This is shown by a decline in fish obtained from the marine environment .
In light of this, the Indonesian government has implemented several initiatives to address biodiversity concerns, including establishing rules and strategic planning.
This is undoubtedly a practice that is likewise carried out globally.
Nevertheless, collaboration among government, scholars, and society remains necessary .
Given the immense expanse of the oceans and the scarcity of scientists, it is essential to establish suitable strategies for effective planning.
The diverse capabilities of each person within the community provide a challenge in assessing the presence of damaged coral reefs or endangered biota .
In addition, the color quality of underwater images is compromised due to color distortion caused by light scattering and changes in object color, resulting in blurry images where coral reef health levels cannot be accurately identified .
, .
Several researchers from different areas of study have made efforts to contribute to the works .
, .
Ae.
Various research studies have distinct areas of investigation, and researchers use augmentation methods to improve performance .
Ae.
, .
, .
Priya and Muruganantham .
employed handcrafted feature descriptors and deep features for coral species classification, utilizing KNN and Random Forest algorithms on the EILAT dataset.
subsequent research .
, they explored the LICWM pattern for coral reef classification, integrating VGG-16 and traditional classifiers like KNN and Random Forest.
LICWMP achieved 98.
8% accuracy combined with VGG-16 on both EILAT and RSMAS datasets, while achieving 98.
on EILAT and 99% on RSMAS when used alone.
Awalludin et al.
combined LBP features extraction with HSV and NN classification to achieve a 92.
60% accuracy in classifying 800 underwater images into four categories.
Thamarai and II.
MATERIALS AND METHOD
Several research studies on the classification of coral reef health indicate that the effectiveness of classification approaches is impacted by the specific qualities of the data or features used and the particular kind of classification The classification method's performance indicates that integrating color, shape, and texture data does not consistently result in poor accuracy.
The accuracy of these features' performance may be altered by using image data processing, mainly by applying image processing before the feature extraction and classification procedures.
The work proposes a coral reef classification framework using the data obtained from Kaggle.
These datasets evaluate the capability to choose augmentation methods using the PSO approach.
The PSO approach employs an adaptive update of inertia weights to enhance performance.
Figure 1 depicts the proposed research, in which the PSO optimization strategy automatically chooses the most suitable mix of augmentation The supplementary datasets that enhance the accuracy of the CNN method's performance in evaluating the fitness function.
Image Augmentation Previous research has employed augmentation methods to enhance the effectiveness of specific classification frameworks .
Ae.
, .
Typically, the augmentation methods include horizontal flipping, random cropping, and color space augmentations.
This research used 58 augmentation methods derived from the Python package "imgaug" .
This library offers various image-processing to process images.
only a subset was used in this research.
The use of different augmentation methods is seen in Table 1.
Selection of the best augmentation method from Table 1 is done using enhanced PSO method (AIWPSO).
The combination of the five best augmentation methods from PSO is chosen from all available augmentation methods.
TABLE I
IMAGE AUGMENTATION METHODS
Name Name Gaussian Blur Average Blur Median Blur Bilateral Blur Motion Blur Multiply Brightness Add To Brightness With Hue And Saturation Multiply Hue And Saturation Multiply Hue Solarize Posterize Equalize Auto Contrast Enhance Color Enhance Contrast Enhance Brightness Multiply Saturation Add To Hue And Saturation Add To Hue Add To Saturation Grayscale Kmeans Color Quantization Uniform Color Quantization Uniform Color Quantization To Nbits Posterize Gamma Contrast Sigmoid Contrast Log Contrast Linear Contrast Sigmoid Contrast Edge Detect Directed Edge Detect Canny Horizontal Flip Vertical Flip Fliplr Flipud Perspective Transform Elastic Transformation Rotation 90 All Channels Clahe Clahe All Channels He Histogram Equalization Sharpen Emboss Enhance Sharpness Filter blur Filter Smooth Filter Smooth More Filter Edge Enhance Filter Edge Enhance More Filter Find Edges Filter Contour Filter Emboss Filter Sharpen Filter Detail Fig.
1 Proposed Method Coralreef dataset The study utilized open-access datasets, including one Kaggle .
ttps://w.
com/datasets/vencerlanz09/healthyand-bleached-corals-image-classificatio.
This dataset was specifically curated to differentiate between healthy and bleached corals, with the images categorized into two groups:
healthy corals .
and bleached corals .
as seen in Fig.
To ensure consistency, all images were resized to a maximum dimension of 300 pixels, maintaining uniformity throughout the dataset.
Augmentation Selection Method based on Adaptive Inertia Weight PSO This research seeks to use Particle Swarm Optimization (PSO) to choose the most effective augmentation strategy.
PSO can choose the most practical combination of augmentation strategies to achieve exceptional accuracy performance .
, .
However, the inertia weight coefficient .
on PSO affects the algorithm's performance .
The inertial weight coefficient value may be adjusted, maintained constant, or elevated when an improved position Fig.
2 Healthy .
and Bleached .
Corals Image Classification Dataset is discovered to ensure the particle remains inside the search space .
This research effectively generates optimum combinations for the training process by applying each parameter of the PSO algorithm.
The PSO phase is shown as Terminate the procedure if a satisfactory solution meets the required minimal error is discovered or the maximum number of iterations is attained.
otherwise, repeat steps .
once Convolution Neural Network as an Evaluation Method The PSO approach was selected when selecting the optimal augmentation method for the most favorable performance This study utilizes a CNN to calculate fitness function values for PSO using accuracy performance A CNN is a mathematical model that employs a parametric design, resembling a neural network approach .
, .
The neural network comprises an input layer, many hidden layers, and an output layer.
The hidden layers are connected to the input layers with configurable weights, and each layer progressively represents increasingly sophisticated aspects of the input images .
The traditional CNN framework comprises a straightforward configuration of modules, including several levels within the CNN architecture, such as convolutional, pooling, and fully linked layers .
Every module transforms the initial input representation at a certain level to a higher, more conceptual level.
The CNN method is shown in Fig.
Estimate the particle swarm and velocity based on the location of each particle, which initially has a random value of zero.
Determine the fitness function used to evaluate each particle's fitness value based on the optimization problem's A fitness function is a quantitative metric used to assess the precision of a specific solution.
In this research, fitness performance was assessed by using classification techniques, especially utilizing CNN.
CNN is applied based on the selected combination of augmentation methods based on particle values.
By comparing the fitness function value of each particle, which is determined from the accuracy value, with the pbest fitness value, we may assess their relative If the current value surpasses the previous highest .
, the current fitness value is updated to match In addition, the value of pbest remains constant.
Calculate the optimal fitness value in a population by considering the fitness values of all particles.
Subsequently, assess the numerical value and compare it with the present global best value.
If the new value is superior, update the global best .
with this value.
otherwise, maintain gbest .
Determine the value of the parameter , which adjusts the inertia weight .
and enables it to adapt based on the isa Meanwhile, individual search ability .
pertains to evaluating the absolute value of the disparity in particle location and the nearby solution.
Fig.
3 CNN Model as an Evaluation Method During identification, higher-level representations enhance the importance of helpful input traits for categorization and diminish irrelevant variations.
CNN predicts fundamental characteristics of pictures.
Comparing the CNN to a regular feedforward neural network reveals that the CNN has a notably lower quantity of connections and parameters.
The presence of these properties facilitates the training process of CNNs.
Nevertheless, it is essential to acknowledge that the optimal performance of a CNN is expected to be somewhat worse in theory.
The capacity of the CNN may be modified by altering its width and depth.
Equations .
govern the determination of the velocity and position update for each particle.
Every particle is defined by a location vector and a velocity vector.
Create a collective group of entities in various dimensions, each being the same size as the feature.
The variable denotes an inertial weight that spans from 0 to 1.
It is an acceleration weight applied to both the particle and the swarm, with a value between 0 and 2.
It is determined by two random numbers evenly distributed throughout the range of 0 to 1.
The particle's location, as is its speed, is confined to a specific .
Convolutional Neural Network model:
The more complex representation in the coral reef categorization method enhances the inclusion of important input information while minimizing irrelevant variations.
The CNN model that has been created accepts RGB pictures with a size of 128 by 128 pixels and three-color channels.
The implementation employs a series of consecutive functions, using an architectural framework that progressively adds Conv2D and MaxPooling2D layers.
The first Conv2D layer of the model has 32 filters, each with dimensions of .
, .
and employs a Rectified Linear Unit (ReLU) activation Subsequently, the MaxPooling2D layer with a pooling size of .
, .
diminishes the spatial dimensions.
The procedure is iterated four times, using Conv2D layers with cumulative filters of 64, 128, 128, and 256.
After each Conv2D layer, there is a MaxPooling2D layer.
where ! represents the !-th iteration in the search process, "OO$ signifies the "-th dimension in the search space, and stands for the inertia weight.
c and c are acceleration constants, while r and r are random values uniformly distributed in the range .
p() and p*) denote the elements of best and best in the "-th dimension, respectively.
The convolution and pooling layers provide an output that is subsequently transformed into a flattened structure using a flattened layer.
The model includes a dense layer of 512 neurons and employs the Rectified Linear Unit (ReLU) activation function.
Ultimately, a Dense layer is appended to the output, with neurons corresponding to the number of classes .
, and a softmax activation function is used to oversee the classification procedure.
Therefore, these attributes of the CNN structure enhance the model's capacity to extract relevant features from RGB pictures .
, 128, .
within the framework of a specific classification assignment.
The model used is shown in Fig.
widely used architectures.
CNN parameters are fine-tuned throughout the process to optimize performance, with each epoch's results meticulously analyzed.
A modified PSO algorithm is employed to identify the most practical combination of five augmentation methods, with weights dynamically adjusted to improve algorithm performance.
Comparative analysis of optimization methods is conducted to yield optimal outcomes.
Training and validation accuracies are meticulously recorded for thorough analysis.
The performance of Adaptive Inertia Weight PSO (AIWPSO) was compared with other optimization algorithms, including the Genetic Algorithm (GA).
Gray Wolf Optimizer (GWO) .
Particle Swarm Optimization (PSO) .
, and Fox Optimization Algorithm (FOX) .
The GA and original PSO methods were chosen as benchmarks because they are commonly used in optimization, while GWO and FOX represent newer swarm-based methods.
Experiments are conducted on machines meeting specific hardware requirements, including an Intel i7-8700 processor and 32GB of RAM, operating on the Windows 11 platform.
Fig.
4 Model CNN i.
RESULTS AND DISCUSSION
Rectified Linear Unit (ReLU):
ReLU activation functions are often used in neural network topologies since they are straightforward and computationally efficient, contributing to their increasing popularity .
ReLU's primary benefit is its capacity to address the "vanishing gradient" issue often seen in other activation functions like sigmoid or tanh.
This enhances efficiency and expedites the model's training process, positively impacting the pace at which convergence occurs.
The ReLU equation is shown in .
,- 0,
Despite the benefits of ReLU, it is crucial to consider its many drawbacks.
Neurons exhibiting harmful activity are excluded from the training process, leading to "dead neurons" that do not contribute to the learning process.
Furthermore, the potential occurrence of "dying ReLU" exists when the neuron stays dormant due to consistently generating a zero This may impede the propagation of gradients and hinder the learning process in these neurons.
Hence, although ReLU might be a viable choice, it requires specific comprehension and management to address any issues that may occur during its use in building neural network models.
Results The first test was conducted to test the epoch parameters, and experiments were carried out to evaluate epochs from epoch 1 to epoch 10.
Figure 5 shows the accuracy performance results based on the epoch parameters used.
The results from Fig.
5 are the Training and Validation accuracy results with the CNN method, which are measured in The best performance results were obtained from pair groups 28, 50, 52, 3, and 4: Sharpen.
Smooth Filter.
Edge Enhance Filter.
Median Blur, and Bilateral Blur.
The augmentation results show details of the original image.
However, from the ten epochs carried out, it turns out that the most frequently chosen augmentation methods 3, 49, and 53, namely Median Blur.
Filter blur, and Filter Edge enhancement, which got the best accuracy values .
hown in Fig.
This is possible because some of these filter methods do not robustly remove image characteristics.
Performance Evaluation The classification algorithms' performance is assessed via the computation of accuracy.
Accuracy is the precise and accurate categorization of all acquired data .
The accuracy values are calculated using equation .
, where t is the number of adequately identified sample data and n represents the total sample data.
- 0 - 1
!/3
Original Sharpen .
Filter Smooth .
Filter Edge Enhance .
Median Blur .
Bilateral Blur .
Fig.
5 Combination of augmentation method from best performance Design Experiment This research emphasizes the critical role of data augmentation in maximizing the efficiency of CNN models.
The study aims to leverage public datasets as the primary resource and introduce a novel approach to enhancing CNN performance through augmentation methods, employing Filter blur .
Median Blur .
Bilateral Blur .
Fig.
6 Most frequently chosen augmentations The highest accuracy performance results were shown when using the 10th epoch, with an accuracy value of 95.
The average training accuracy performance for ten epochs is Descriptively, it is shown that the training accuracy values from Fig.
7 range between 91.
43% and 95.
64%, while the validation accuracy values range between 80.
87% and The average training accuracy is around 93.
while the average validation accuracy is around 82.
relatively low standard deviation indicates that accuracy values are within a limited range of the average.
Positive skewness in validation accuracy suggests that the data distribution is slightly skewed to the right, meaning there are more values on the right side of the mean.
The maximum and minimum values reflect the highest and lowest accuracy achieved in the test, respectively.
From this explanation, it can be seen that the training performance results on CNN provide higher performance than the accuracy results in the validation Indications of fluctuations in the results show an increasing trend in the training accuracy results.
In contrast, the validation accuracy values show stagnation, which makes it possible that there is still overfitting of the resulting performance results even though the difference is Especially when seen from the average accuracy performance obtained from each training epoch.
96%, followed by AIWPSO with 95.
64% and GWO Nonetheless, the average and maximum values of all algorithms show relatively comparable performance, highlighting the strong ability of each algorithm to find optimal solutions.
TABLE II
COMPARISON OF OPTIMIZATION METHODS FOR SELECTING THE BEST
AUGMENTATION METHOD (IN %)
Epoch
AVG
AIWPSO
GWO
PSO
FOX
Fig.
8 Average accuracy of optimization performance The results of statistical analysis from Table 2 include several essential parameters, which include average, standard error, median, standard deviation, sample variance, kurtosis, skewness, range, minimum and maximum values, and the sum of the performance of each optimization method is based on data from Table 2.
From the results of this analysis, the average AIWPSO performance is 93.
44% with a standard error of 0.
42, a median of 93.
82%, and a range of performance between the minimum value of 91.
43% and a maximum of In this case, the confidence level is 95%, which provides a confidence interval for the average performance of each optimization method.
These findings provide a deeper understanding of the characteristics and performance variations of each optimization method evaluated, which can be used to select the most appropriate method for solving a given optimization problem.
From the statistical analysis results based on testing using Friedman, a p-value of 0.
0171 was obtained based on testing each epoch value in optimization.
The results show the statistical significance of the differences between each epoch in the optimization method.
However, the p-value obtained from testing based on the performance results of each optimization method is 0.
7358, indicating no significant difference between the optimization results obtained.
This is supported by the relatively comparable average ranges given .
an be seen in Fig.
The results of the optimization method in Table 2 are the best accuracy performance obtained based on the selection of the augmentation method shown in Table 3.
The numbers in Table 3 are the augmentation methods used according to the research plan in Table 1.
The data results in Table 3 illustrate Fig.
7 Accuracy proposed method This analysis provides the results of the Wilcoxon test regarding the difference between training and validation The results of Fig.
1 for each epoch show that in this analysis, two estimators of mean disagreement were used:
the Binomial estimator and the Hodges-Lehmann estimator.
The binomial estimator shows a mean difference of -10.
while the Hodges-Lehmann estimator shows a mean difference of -10.
From these results, it can be concluded that there is a significant difference between training accuracy and model validation accuracy.
The p-value obtained from the Wilcoxon test is 0.
0020, below the commonly used alpha level .
, so we can reject the null hypothesis that there is no difference between training and validation accuracy for each epoch.
Comparison based on other Optimization Methods In this research, the performance of five optimization algorithms was evaluated: i.
AIWPSO.
GA.
GWO.
Original PSO, and FOX, in solving a particular problem for ten epochs and ten population for the number of swarms in PSO.
GA.
GWO or the number of chromosomes for GA.
The test results in Table 2 show variations in the performance of each Regarding average performance scores.
AIWPSO and GA stood out with scores of 93.
44% and 93.
50%, while GWO recorded a score of 93.
Meanwhile, in the maximum performance value.
GA achieved the highest score the augmentation method chosen for each optimization algorithm, namely AIWPSO.
GA.
GWO.
PSO, and FOX.
The selection of the augmentation method is based on each optimization algorithm's characteristics to achieve accurate performance results.
For example.
AIWPSO uses sharpening techniques to improve image sharpness, while GA chooses histogram equalization to flatten the pixel intensity optimal solution.
Statistical analysis also highlights the importance of selecting an augmentation method that suits the characteristics of each optimization algorithm to improve classification performance.
For example, the use of sharpening techniques in AIWPSO and histogram equalization in GA are related to the characteristics and objectives of each algorithm.
Thus, choosing the proper augmentation method can help improve input data quality and variety, thereby increasing classification data accuracy.
TABLE i THE RESULTS OF SELECTING THE AUGMENTATION METHOD ARE BASED ON IV.
CONCLUSION
THE OPTIMIZATION METHOD
Combination AIWPSO
GWO
PSO
FOX
GWO, on the other hand, selects the edge to enhance more filters to increase the contrast of the edges of the image, while PSO uses multiple saturation to strengthen the colors in the With its different approach.
FOX chooses elastic transformation to introduce elastic deformation to images to enrich data diversity.
Thus, using appropriate augmentation methods can help improve the performance of classification algorithms in achieving their primary objective, namely achieving the highest accuracy in data processing.
This analysis of augmentation methods provides valuable insight into how different approaches can improve the quality and variety of input data in the context of classification algorithms.
Discussion The results show that testing was carried out on the epoch parameters to evaluate the Convolutional Neural Network (CNN) performance of various augmentation methods.
These results demonstrate that the best performance was obtained in the 10th epoch with an accuracy of 95.
Additionally, there are variations in performance between different augmentation methods, where specific augmentation methods tend to be selected more frequently in the training process, such as Median Blur.
Filter Blur, and Filter Edge Enhancement.
Furthermore, a performance comparison between five optimization algorithms shows that Adaptive Inertia Weight Particle Swarm Optimization (AIWPSO) and Genetic Algorithm (GA) stand out with the highest average The AIWPSO-based augmentation method could provide an average accuracy performance of 93.
44%, below the GA method, namely 93.
There is a difference of 0.
between the AIWPSO and GA methods.
It is also shown that AIWPSO obtained a higher score than GWO.
FOX, and original PSO, namely 93.
34%, 92.
51%, and 93.
The difference in performance results is not more than 1%.
It can be seen that the average ranges obtained are relatively comparable, which is below the results of the Friedman test, which received a p-value of 0.
7358, which means that there is no significant difference resulting from the optimization performance.
This optimization method selects augmentation methods that can provide the highest accuracy performance based on the CNN method.
Nevertheless, the statistical results show no significant difference between the performance of different optimization algorithms.
This indicates that each algorithm has a strong ability to find the This study employs optimization methods to determine the most effective combination of augmentation methods for enhancing image quality.
This combination of augmentation methods serves as an initial preprocessing step before data extraction from the neural network, particularly in addressing the challenge of coral reef classification, especially distinguishing between healthy and unhealthy reefs or those undergoing bleaching.
Evaluation of epoch parameters in Convolutional Neural Network (CNN) methodology for image classification revealed that the highest accuracy was achieved at the 10th epoch, with an average training accuracy 96% and validation accuracy of 82.
Furthermore, analysis was conducted on five different optimization algorithms, with Adaptive Inertia Weight Particle Swarm Optimization (AIWPSO) and Genetic Algorithm (GA) exhibiting the highest average performance.
While there were variations in maximum performance among the algorithms, no statistically significant differences were observed in each method's optimization outcomes.
It has been demonstrated that employing appropriate augmentation methods enhances the quality and diversity of data in image classification.
Consequently, the primary conclusion is that selecting suitable optimization, such as GWO .
and augmentation methods can effectively improve performance in image classification tasks utilizing CNNs.
ACKNOWLEDGMENT
This research is supported by Universitas Dian Nuswantoro, specifically by the Research Center for Intelligent Distributed Surveillance and Security, with a particular emphasis on Artificial Intelligence Studies in Nature Conservation and Natural Disasters.
REFERENCES