JIPK.
Volume 17 No 3 October 2025 Sinta 1 (Decree No: 158/E/KPT/2.
e-ISSN:2528-0759.
p-ISSN:2085-5842 Available online at https://e-journal.
id/JIPK JIPK
(JURNAL ILMIAH PERIKANAN DAN KELAUTAN)
Scientific Journal of Fisheries and Marine Research Article Deep Learning Models Performance on Marine Fish Species Classification Nur Muhammad Afiq Anang , and Ezmahamrul Afreen Awalludin* Faculty of Fisheries and Aquaculture Sciences.
Universiti Malaysia Terengganu, 21030 Kuala Nerus.
Terengganu.
Malaysia.
Abstract
ARTICLE INFO
Received: April 16, 2025 Accepted: July 15, 2025 Published: August 01, 2025 Available online: Sept 27, 2025 *) Corresponding author:
E-mail: e.
afreen@umt.
Keywords:
AlexNet Deep Learning Fish Classification GoogLeNet Residual Network 50 This is an open access article under the CC BY-NC-SA license .
ttps://creativecommons.
org/licenses/by-nc-sa/4.
Identifying marine fish species accurately can be difficult due to their subtle anatomical and color pattern similarities, which often result in misclassification during ecological assessments and fisheries operations.
Manual identification methods are time-consuming and prone to errors especially in high throughput environments such as fish markets.
In this study, transfer learning is used to evaluate three deep learning models ResNet-50.
AlexNet and GoogLeNet on a total of 20,325 images from twenty marine fish species acquired from Kuantan (Pahan.
and Mengabang Telipot (Kuala Neru.
Malaysia.
All images were morphologically classified as complete fish, head, body and tail.
The dataset was subjected to preprocessing procedures which encompassed image resizing, pixel normalization and data augmentation techniques that consists of random rotation (A15A), horizontal flipping, adjustments to brightness and contrast (A20%) and Subsequently, the dataset was partitioned into 80% training set .
,260 image.
, 10% validation set .
,032 image.
and 10% testing set .
,033 image.
The classification patterns were analyzed using confusion matrices and standard metrics such as accuracy, precision and recall.
ResNet-50 outperformed other models achieving ideal results with 100% accuracy, precision and recall in every category.
With 99.
5% and 99.
4% accuracy.
GoogleNet and AlexNet came in second and third, respectively.
This study shows that deep learning models especially ResNet-50 achieved an accurate and efficient way to classify fish species automatically.
With multi-view images, data augmentation and transfer learning, the model performs well even in difficult visual conditions.
These results support its use in real-time fisheries monitoring, biodiversity studies, and environmental impact assessments.
Cite this as: Anang.
, & Awalludin.
Deep Learning Models Performance On Marine Fish Species Classification.
Jurnal Ilmiah Perikanan dan Kelautan, 17.
:608Ae626.
http://doi.
org/10.
20473/jipk.
Copyright A2025 Faculty of Fisheries and Marine Universitas Airlangga Anang et al.
/ JIPK, 17.
:608-626
Introduction Fish classification plays a vital role in ecological monitoring, biodiversity conservation, and sustainable fisheries management (Yang et al.
, 2.
Traditionally, this process has relied on morphological characteristics and DNA barcoding techniques.
While these methods provide results, they demand significant resources and require highly skilled personnel, limiting their practicality for large-scale applications (Iqbal et al.
, 2021.
Hilal et al.
, 2.
Moreover, the hands-on aspect of these methods makes them susceptible to human mistakes and personal bias, particularly when addressing uncommon or morphologically akin Consequently, the scalability and practicality in contemporary fisheries science continue to pose significant challenges.
Recently, deep learning an advanced branch of artificial intelligence has emerged as a powerful tool for automating fish classification, offering high accuracy with minimal human intervention (Saleh et , 2022.
Hasan et al.
, 2022.
Han et al.
, 2.
Deep learning models, particularly Convolutional Neural Networks (CNN.
have significantly transformed the field of image recognition by effectively extracting complex hierarchical features directly from raw image data (Peddina and Mandava, 2025.
Prasetyo et al.
Zhang et al.
, 2022.
Wang et al.
, 2023.
Iqtait et , 2.
In contrast to traditional computer vision methods that rely on manual feature extraction.
CNNs autonomously learn spatial and temporal patterns, thereby improving accuracy and reducing the necessity for specialized domain knowledge.
Nevertheless, the capacity of CNNs to classify various components of fish anatomy, including the head, body, and tail, remains challenging due to high visual similarity in texture and shapes across species, which is less distinct than in typically general objects classification (Lan et , 2020.
Ahmad et al.
, 2.
A typical CNN model is made up of several essential elements: the input image, convolutional layer, pooling layer, activation function, fully connected layer, and output layer (Rawat and Wang, 2017.
Song et al.
, 2.
The process starts with the input image, which represents visual data from the pixels of a digital image.
The convolutional layer employs a kernel or filter to automatically identify specific features, producing a comprehensive channel representation of the input images.
Research indicates that CNN models like ResNet.
AlexNet, and GoogLeNet effectively classify marine organisms, including various fish species, plankton, and coral (Allken et al.
, 2019.
Wang et , 2021.
Veluswami and Panneerselvam, 2021.
Zhou et al.
, 2022.
Alinsug et al.
, 2024.
Zhai et al.
, 2.
However, the growing potential of research reveals significant deficiencies in model evaluation, especially concerning classification accuracy across different fish body areas.
This area remains poorly understood, even though comprehensive categorization is crucial for effective real-time underwater monitoring Furthermore, many current models rely on broad benchmark datasets, which limit their capacity to adapt specific local marine environments.
Limited research has methodically evaluated various CNN architectures on a consistent dataset, employing uniform performance metrics specifically for fish classification (Dong et al.
, 2023.
Rawat and Wang, 2017.
Sun et al.
, 2.
Current studies infrequently investigate the performance of various CNNs across diverse data conditions or with images taken in authentic aquatic settings (French et al.
, 2020.
Ismail et al.
, 2021.
Kaya et al.
, 2.
Additionally, the visual background and camera settings used during image acquisition are often overlooked, even though they significantly impact feature extraction and model Therefore, itAos crucial to conduct a more indepth investigation into fish image classification that considers both full-body images and region-specific segments across different CNN models.
This study examines how effectively three prominent CNN architectures ResNet50.
AlexNet, and GoogLeNet classify 20 local marine fish species using a dataset collected in controlled laboratory settings.
The research aims to address this question on How accurate are CNN models in classifying fish species when employing augmented and region-specific image data? The experimental design includes both full-body images and cropped images that concentrate on the head, body, and tail, aimed at evaluating classification performance in specific regions.
Model performance is evaluated through established metrics like accuracy, precision, and recall, complemented by confusion matrices and training loss curves (Ahmed et al.
, 2.
This research propels the field of ecological AI by addressing several research gaps through innovative dataset design, data augmentation techniques, and a comparative analysis of CNN architectures (Ben Tamou et al.
, 2022.
Deka et al.
, 2.
The study concentrates on developing a carefully annotated fish image dataset, segmented by regions and representative of authentic marine environments (Zheng et al.
, 2.
The study presents a comparison of three notable CNN models, highlighting their real-world uses and limitations.
The results aim to improve scalable and accurate fish classification by offering an advanced AI-based approach to enhance fisheries management, species JIPK: Scientific Journal of Fisheries and Marine JIPK Vol 17 No 3.
October 2025 | Deep Learning Models Performance on Marine Fish Species Classification monitoring, and marine conservation efforts.
To tailor the models for fish species identification, transfer learning was applied (Deka et al.
, 2023.
Zheng et , 2.
Each model underwent training across ten separate sessions, using the same hyperparameters to ensure consistency and evaluate performance through various iterations.
This standardized training approach guaranteed stable results despite the typical random variations associated with deep learning optimization Materials and Methods 1 Material The tools used in this research were selected for their ability to obtain high-resolution aquatic images suitable for classification purposes and to recreate a controlled aquatic environment.
1 The equipments Fish images were acquired using a Huawei P30 Lite smartphone equipped with a 24 MP wide lens and AI scene recognition, ensuring consistent high-resolution output suitable for training deep learning models.
This device features a triple-lens setup comprising a 24 MP wide lens, an 8 MP ultra-wide lens, and a 2 MP depth sensor, all precisely designed to deliver high-resolution images of 1080 y 2312 pixels.
The integration of AI scene recognition and image stabilization features guarantees consistent and sharp image quality, which is crucial for training deep learning models.
Furthermore, the study used a standard glass aquarium measuring 30 cm in height, 30 cm in width, and 40 cm in length as the observation chamber.
The transparent glass walls minimized distortion and maintained uniform lighting throughout the image acquisition process.
The aquarium was designed with standardized dimensions and substrate depth to replicate natural marine conditions while optimizing image clarity and fish mobility for classification purposes (Zheng et al.
, 2.
2 The materials The research initiative concentrated on the classification of fish species utilizing Convolutional Neural Networks (CNN) and depended on two essential materials: sand and saltwater.
These materials were chosen due to their significant role in the marine ecosystem, which is imperative for the study of various fish species.
By using sand and saltwater, researchers were able to create environments that closely mimic the natural habitats of the fish being studied.
Furthermore, the artificial saline mixture of saltwater was meticulously prepared and introduced into the aquarium to a depth of 20 centimeters, yielding an estimated volume of 24 liters, determined by the tankAos dimensions .
cm length y 30 cm width y 30 cm heigh.
The designated water depth was carefully selected to mimic the natural environment of the fish species and to enhance optimal camera focus from different angles.
In addition, a consistent layer of 5.
0 cm thick natural beach sand was evenly distributed along the tankAos bottom.
Adding beach sand achieved two main goals: .
it replicated the fish speciesAo natural habitat, and .
it enhanced the visual backgroundAos complexity, which supports the development of effective computer vision models (Zheng et al.
, 2.
3 Ethical approval This research did not include the use of live experimental animals, humans, or any protected species.
Consequently, there was no requirement for ethical approval from an institutional animal care and use Nevertheless, the research adhered to established ethical principles for gathering image-based data in environmental contexts.
As precaution, any future studies involving biological samples or live subjects will secure ethical clearance in accordance with institutional policies and national regulations.
2 Method 1 Experimental design The experimental framework used in this study and provides a systematic overview of the research workflow.
This framework encompasses data collection, data augmentation, dataset training, model evaluation, and classification test.
These components are essential for comprehending the contributions of each phase to the overall success of the research objectives (Figure .
By carefully following this framework, the study aims to ensure strong results and improve the reliability of the findings.
2 Fish specimen collection Twenty marine fish species were collected for classification, as detailed in Table 1.
Data collection took place over a 12-month period, from January to December 2024, enabling analysis of seasonal variations in species availability.
Specimens were obtained from various local markets and vendors in Kuantan (Pahan.
and Mengabang Telipot.
Kuala Nerus (Terenggan.
Malaysia.
These locations were thoughtfully chosen to ensure a thorough representation of the marine biodiversity along the East Coast of Peninsular Malaysia.
3 Data augmentation Copyright A2025 Faculty of Fisheries and Marine Universitas Airlangga Anang et al.
/ JIPK, 17.
:608-626
This study employed a range of image augmentation techniques on the unprocessed dataset to enhance the robustness and generalizability of the fish classification model.
The methodologies utilized were meticulously developed to improve the dataset and incorporate variability that accurately reflects real-world scenarios in a synthetic environment.
After acquiring images under controlled conditions, each original image was enhanced through various geometric and photometric transformations.
Drawing from the research conducted by Yasin et al.
, this study examined various image manipulation techniques, such as rotations within a range of -30A to 30A, horizontal and vertical flips, scaling, cropping, and modifications to brightness and contrast.
The selected methodologies endeavour to replicate the natural orientation of fish as well as the variations in lighting.
This approach facilitates the modelAos ability to acquire invariant features across a range of conditions (Li et al.
, 2.
The execution of these transformations was particularly imperative for species with limited initial sample sizes, thereby ensuring sufficient representation throughout the dataset.
Images were captured using a stationary camera setup that ensured consistent lighting, maintained a 90-degree angle from above, and was positioned 15 cm away, which promoted uniformity among samples.
This method was designed to minimize environmental noise and facilitate precise feature extraction during model training and validation.
4 Preparation of the training dataset After augmentation, the dataset was divided into training and testing subsets with an 80:20 stratified split to maintain the proportional representation of all species in both sets.
Before training began, each image was resized to 224 y 224 pixels to comply with the input specifications mandated by the convolutional neural network (CNN).
Taxonomists performed extensive validation and label annotation on each image, ensuring ground truth accuracy.
The research employed three advanced pre-trained CNN models ResNet50.
AlexNet, and GoogLeNet chosen for their established image classification effectiveness and capacity for fine-tuning on specific datasets.
Transfer learning was Figure 1.
The process of experimental design.
The operations resulted in an expansion of the dataset to 20,325 images, which showcased 20 distinct fish species.
This enhanced dataset was specifically designed to address class imbalance and to enrich the variety of included visual features.
Considerable efforts were undertaken to ensure that augmentation was implemented consistently across all classes, thereby preserving the balance of the dataset and mitigating the emergence of any potential bias.
This approach aligns with the best practices noted in recent research (Okafor et al.
, 2.
, indicating that augmentation improves classification performance and acts as a regularization technique during model training.
To develop a varied image dataset for model training, images of each species were captured in two controlled settings: .
on a flat surface with laminated white A3 paper as the background and .
in an artificial underwater environment, as shown in Figures 2 and Figure 3.
The underwater scene included a transparent aquarium filled with saltwater and natural beach sand from local coastal areas to replicate genuine marine conditions.
This two pronged background strategy sought to improve the modelAos versatility by integrating both controlled and naturalistic imaging employed to tailor the models for the specific task of recognizing fish species.
Each model underwent ten training sessions, utilizing the same hyperparameters to ensure consistency and evaluate performance across various iterations.
This consistent training methodology guaranteed that the outcomes were not affected by the random fluctuations typically associated with deep learning optimization techniques.
Furthermore, to assess the distinguishing capability of various anatomical areas, classification was conducted utilizing both full-body images and segmented images that concentrated on the fish head, mid-body, and tail.
This comprehensive analysis across multiple regions sought to pinpoint the body part that holds the most significant visual cues for precise species classification, providing essential insights for the advancement of more effective and localized recognition systems.
5 Classification and assessment After finishing the model training, evaluations for classification were carried out on all three CNN Performance evaluation was conducted through three main tools: the confusion matrix, training progress graphs, and classification test reports.
The JIPK: Scientific Journal of Fisheries and Marine JIPK Vol 17 No 3.
October 2025 | Deep Learning Models Performance on Marine Fish Species Classification .
Figure 2.
Images of fish species presented on a laminated A3 background as well as under underwater conditions: .
Abalistes stellatus, .
Alectis indica/indicus, .
Carangoides gymnosthetus, .
Pampus argenteus, .
Epinephelus bleekeri, .
Euthynnus affinis, .
Johnius dussumieri, .
Lates calcarifer, .
Lethrinus lentjan, .
Lutjanus gibbus, .
Megalaspis cordyla, .
Nemipterus furcosus, .
Pampus argenteus, .
Parastromateus niger, .
Psettodes erumeiA .
Scarus ghobban, .
Selar crumenophthalmus, .
Seriolina nigrofasciata, .
Trachinotus blochii, .
Trichiurus lepturus.
Copyright A2025 Faculty of Fisheries and Marine Universitas Airlangga Anang et al.
/ JIPK, 17.
:608-626
confusion matrix facilitated the calculation of fundamental performance metrics, encompassing overall accuracy, precision, and recall.
Accuracy is defined as the ratio of correctly classified samples to the total number of samples, whereas precision assesses the proportion of true positives among all positive predictions.
Recall, also known as sensitivity, indicated the percentage of true positives accurately recognized by the model.
The metrics were derived using the established formula.
Accuracy is calculated by taking the sum of true positives and true negatives, divided by the total of true positives, false positives, true negatives, and false Accuracy = (Tp T.
/(Tp Tn Fp F.
(Eq .
Accuracy is determined by the ratio of true positives to the sum of true positives and false positives.
Precision = TP/(Tp FP) x100%.
(Eq .
Recall is calculated as the ratio of true positives to the sum of true positives and false negatives.
Recall = Tp/(Tp F.
(Eq .
Where true positives (T.
, true negatives (T.
, false positives (F.
, and false negatives (F.
of the modelAos predictions.
These metrics provided a comprehensive view of model performance, enabling a detailed analysis that goes beyond mere accuracy numbers.
Graphs illustrating training progress were created to monitor the changes in loss and accuracy throughout the training epochs.
The visualizations played a crucial role in evaluating the learning behaviour of each CNN and pinpointing possible challenges like under fitting or Given the constraints of the hardware, we presented representative graphs for each model to illustrate the trends in convergence.
The classification tests involved comparing results across various anatomical regions to determine which segment yielded the highest accuracy.
This segmentation method supports the overarching goal of creating optimised, resource-efficient models for immediate application in aquaculture systems.
3 Analysis Data The evaluation of model performance was conducted comprehensively through both statistical and visual techniques utilizing MATLAB Software 2022b, which is licensed to Universiti Malaysia Terengganu.
Confusion matrices were created for each model to as- sess the accuracy of predictions for each fish species.
The matrices offered valuable insights into particular inter-class misclassifications, especially among visually similar species, aiding in pinpointing the areas where the model faced challenges in distinguishing between different fish types.
Class-wise analysis of precision and recall scores was conducted to evaluate the performance of each model in managing both prevalent and infrequent species.
Classification reports compiled all performance metrics and were utilized to evaluate models comprehensively, allowing for the identification of the most effective architecture.
The examination further encompassed reviewing training progress charts for each CNN and graphing training accuracy and loss over the epochs.
The curves facilitated the evaluation of how the model converges and helped pinpoint any irregularities in the training The study aimed to assess not only model accuracy but also the practicality of implementing AI-driven classification within actual aquaculture and marine monitoring settings.
The research offered important insights into practical deployment considerations by analysing how effectively the models distinguish species based on various anatomical features.
This analysis highlighted the importance of image quality control, camera positioning, and habitat variability.
The data indicated that convolutional neural network-based models can attain impressive accuracy in species classification if trained with well-augmented and balanced datasets in controlled settings.
These findings contribute to the advancement of automated fish identification systems designed to support sustainable fisheries and ecological monitoring efforts.
Results and Discussion 1 Result The confusion matrix serves as a tabular representation employed to assess the performance of a classification model by juxtaposing its predicted labels with the actual labels.
In the context of this study, it elucidates the precision with which the deep learning model differentiates each species of marine fish.
The matrix exhibits precise predictions along the diagonal.
however, the off-diagonal entries signify instances of Through a comprehensive analysis of this matrix, it can differentiate specific species that the model consistently misclassifies, thereby enabling subsequent improvements.
The graph depicting training progress effectively visualizes the modelAos learning trajectory gradually, typically illustrating metrics such as accuracy and JIPK: Scientific Journal of Fisheries and Marine JIPK Vol 17 No 3.
October 2025 | Deep Learning Models Performance on Marine Fish Species Classification loss throughout the training process epochs.
In this research, the graphical representation demonstrates the enhancement of the modelAos performance across each epoch.
A consistent rise in accuracy accompanied by a corresponding decline in loss indicates effective Conversely, any irregularities may signal complications such as overfitting or under fitting.
is imperative to monitor this graph to ensure that the model is training as anticipated and to facilitate the implementation of necessary adjustments.
Figure 4 displays the confusion matrix results from the initial training sessions of each Convolutional Neural Network (CNN) model.
In Figure 4.
, the ResNet50 modelAos confusion matrix shows outstanding performance, achieving an accuracy, precision, and recall of 100%.
This result demonstrates that the ResNet50 model effectively identified all fish species without any false positives or negatives.
Figure 4.
presents the confusion matrix for the AlexNet model, which recorded an accuracy of 99.
9%, alongside Table 1.
Taxonomy of twenty selected fish species in this study No.
Family Genus Species Balistidae Abalistes Abalistes stellatus Carangidae Scyris Alectis indica/indicus Carangidae Carangoides Carangoides gymnosthetus Serranidae Epinephelus Epinephelus areolatus Serranidae Epinephelus Epinephelus bleekeri Scombridae Euthynnus Euthynnus affinis Sciaenidae Johnius Johnius dussumieri Latidae Lates Lates calcarifer Lethrinidae Lethrinus Lethrinus lentjan Lutjanidae Lutjanus Lutjanus gibbus Carangidae Megalaspis Megalaspis cordyla Nemipteridae Nemipterus Nemipterus furcosus Stromateidae Pampus Pampus argenteus Carangidae Parastromateus Parastromateus niger Psettodidae Psettodes Psettodes erumei Scaridae Scarus Scarus ghobban Carangidae Selar Selar crumenophthalmus Carangidae Seriolina Seriolina nigrofasciata Carangidae Trachinotus Trachinotus blochii Trichiuridae Trichiurus Trichiurus lepturus 1 Confusion table perfect precision and recall at 100%.
While its accuracy Copyright A2025 Faculty of Fisheries and Marine Universitas Airlangga Anang et al.
/ JIPK, 17.
:608-626
Table 2.
The process of data augmentation applied to enhance the training dataset for deep learning models Pre-Processing Whole Picture Brightness/Contrast .
harpen 25%):
0%/0% .
10%/10% .
20%/20% .
30%/30% Images .
Same Brightness/Contrast as above images (Sharpen -25%): .
, .
, .
, .
Head (Croppe.
Brightness/Contrast:
0%/0% .
10%/10% .
20%/20% .
30%/30% .
Same Brightness/Contrast as above images (Sharpen -25%): .
, .
, .
, .
Body (Croppe.
Brightness/Contrast:
0%/0% .
10%/10% .
20%/20% .
30%/30% Same Brightness/Contrast as above images (Sharpen -25%): .
, .
, .
, .
JIPK: Scientific Journal of Fisheries and Marine JIPK Vol 17 No 3.
October 2025 | Deep Learning Models Performance on Marine Fish Species Classification Tail (Croppe.
Brightness/Contrast:
0%/0% .
10%/10% .
20%/20% .
30%/30% .
Same Brightness/Contrast as above images (Sharpen -25%): .
, .
, .
, .
Rotation .
270o Same rotation as above images (Sharpen -25%): .
, .
, .
, .
*The underwater images also through a similar data augmentation process as above.
Aquarium Water Salt Water = 20.
0 cm Beach Sand = 5.
0 cm Figure 3.
The aquarium has been thoughtfully organized to replicate an underwater ecosystem, thereby enabling the acquisition of authentic images of fish.
Copyright A2025 Faculty of Fisheries and Marine Universitas Airlangga Anang et al.
/ JIPK, 17.
:608-626
is slightly lower than that of ResNet50, its flawless precision and recall illustrate its capability in classifying fish species in this training period.
Figure 4.
shows the confusion matrix for the GoogLeNet model, which reached an accuracy of 99.
6%, with perfect precision and recall .
%).
Although its accuracy is a bit lower than that of AlexNet and ResNet50.
GoogLeNet still exhibited strong performance in the classification task, resulting in no false positives or negatives.
Table 3 summarizes the confusion matrix results for each model and training session, enabling a thorough comparison of performance across different models.
The results highlight the effectiveness of the Convolutional Neural Network (CNN) models in classifying fish species, with ResNet50 noted as the model demonstrating the best performance efficiency.
The discrepancies in performance levels among these models may be ascribed to their respective architectural intricacies, particularly the utilization of deeper layers and skip connections in ResNet50, which contribute to its enhanced accuracy.
Although AlexNet and GoogLeNet display marginally lower accuracy rates, both models demonstrate commendable performance, achieving flawless precision and recall metrics, thereby establishing them as dependable alternatives for this classification task.
for the various CNN models: .
ResNet50, .
AlexNet, and .
GoogLeNet.
Due to hardware limitations, specifically the constraints imposed by a single-core CPU, training was executed solely once for each model.
The dataset, comprising 20,325 images, posed challenges for the system in terms of repeated processing.
Each graph illustrates the validation value, which serves as an indicator of the modelAos performance concerning accuracy following each training session.
The loss value denotes the number of incorrect predictions made during the training process, functioning as a critical metric for assessing the modelAos learning progression over time.
A reduced loss value signifies superior model performance, whereas an elevated value indicates opportunities for enhancement.
The recorded training time encompasses the total duration taken for the model to complete the training, including the time required for data loading.
Progress in training was monitored subsequent to each designated epoch, with live updates reflecting the evolution of various parameters throughout the session.
This approach enabled real-time tracking of model performance.
Nonetheless, hardware limitations restricted us to a single training session for each model.
The training was limited to two epochs, using sixty-four mini batches and a learning rate of 0.
001 to enhance the modelAos Table 3.
Comparison of classification performance through confusion matrices for ResNet-50.
AlexNet, and GoogleNet ResNet50 Training AlexNet GoogLeNet Average 2 Training progress graph Figure 5 presents the training progress graphs accuracy, considering the limited computational resources.
In future projects, we might leverage more advanced hardware, allowing for additional training JIPK: Scientific Journal of Fisheries and Marine JIPK Vol 17 No 3.
October 2025 | Deep Learning Models Performance on Marine Fish Species Classification sessions and more significant potential improvements to these models.
study, we selected 10 fish species from a pool of 20 distinct species, ensuring that new images were used Table 4.
The classification test on 10 fish species Classification Model ResNet50 AlexNet GoogLeNet Training *This compiles results including for whole part of the fish, head, body, and tail .
imilar result.
3 Classification test The classification test was conducted after each training session to validate the results shown in the confusion matrix, as presented in Table 4.
For this for testing, which were excluded from the training dataset to reduce bias.
The tested species included Az Pampus argenteus.
Psettodes erumei.
Scarus ghobban, and Trachinotus blochii.
Copyright A2025 Faculty of Fisheries and Marine Universitas Airlangga Anang et al.
/ JIPK, 17.
:608-626
JIPK: Scientific Journal of Fisheries and Marine JIPK Vol 17 No 3.
October 2025 | Deep Learning Models Performance on Marine Fish Species Classification Figure 4.
The confusion matrix table in training 1: (A) ResNet50, (B) AlexNet, (C) GoogLeNet Training Progress .
-Dec-2024 21:04:.
Result Validation accuray:
Training finished:
Max epochs completed Training Time Start time 18-Dec-2024 21:04:43 Elapsed time:
1346 min 27 sec Training Cycle Epoch:
2 of 2- Iteration 444 0f 444 Iteration per epoch:
Maximum iterations:
Validation Frequency:
3 iterations Other information Hardware resource:
Single CPU Learning rate schedule Constant Learning rate:
Copyright A2025 Faculty of Fisheries and Marine Universitas Airlangga Anang et al.
/ JIPK, 17.
:608-626
Training Progress .
-Dec-2024 17:42:.
Result Validation accuray:
Training finished:
Max epochs completed Training Time Start time 18-Dec-2024 17:42:28 Elapsed time:
193 min 6 sec Training Cycle Epoch:
2 of 2 Iteration 444 0f 444 Iteration per epoch:
Maximum iterations:
Validation Frequency:
3 iterations Other information Hardware resource:
Single CPU Learning rate schedule Constant Learning rate:
Training Progress .
-Dec-2024 14:05:.
Result Validation accuray:
Training finished:
Max epochs completed Training Time Start time 18-Dec-2024 14:05:36 Elapsed time:
191 min 25 sec Training Cycle Epoch:
2 of 2 Iteration 444 0f 444 Iteration per epoch:
Maximum iterations:
Validation Frequency:
3 iterations Other information Hardware resource:
Single CPU Learning rate schedule Constant Learning rate:
Figure 5.
Training progress graphs for deep learning models: (A) ResNet-50, (B) AlexNet, and (C) GoogLeNet.
JIPK: Scientific Journal of Fisheries and Marine JIPK Vol 17 No 3.
October 2025 | Deep Learning Models Performance on Marine Fish Species Classification Table 5.
Accuracy comparison from confusion matrix table and training progress graph on similar
models in other studies CNN Models Studies ResNet50
AlexNet
GoogLeNet
CMT
TPG
CMT
TPG
CMT
TPG
(Deka et al.
, 2.
(Zhou et al.
, 2.
(Zhou et al.
, 2.
(Rauf et al.
, 2.
Proposed Study *CMT= Confusion Matrix Table.
TPG= Training Progress Graph The accuracy results derived from the confusion matrix indicate that ResNet 50 attained a perfect accuracy of 100%, whereas AlexNet achieved an accuracy of 99.
4%, and GoogLeNet reached an accuracy ResNet 50 maintained consistent performance, while AlexNet reached a peak training accuracy of 100% in sessions 2, 5, 6, and 8, despite fluctuations during these sessions.
Although AlexNet did not match ResNet 50Aos reliability, its simpler architecture with fewer layers produced noteworthy results for a Convolutional Neural Network (CNN), mainly because it was one of the pioneering architectures for deep learning tasks.
GoogLeNet also produced strong results, achieving a maximum accuracy of 99.
which occurred during training sessions 3, 6, and 8.
As illustrated in Table 5, the Training Progress Graph ResNet 50 consistently exhibited the best performance with 100% validation accuracy.
In contrast.
AlexNet and GoogLeNet attained validation accuracies of 97.
9% and 98.
71%, respectively.
ResNet 50 reached 90% accuracy early in the training process .
round iteration .
, while the other models required more time to reach that threshold.
The deeper layers in ResNet 50 allowed it to capture more intricate patterns in the images, thus achieving superior accuracy.
However, adding these deeper layers led to a longer training time for the model.
Specifically.
ResNet 50 needed 1,346 minutes and 27 seconds to finish its training considerably longer than AlexNet, which took 193 minutes and 6 seconds, and GoogLeNet at 191 minutes and 25 seconds.
While the deeper layers of ResNet 50 enhance its accuracy, longer training times can pose difficulties on limited hardware, as high temperatures during prolonged training may negatively affect outcomes.
Nonetheless, all three models exhibited strong performance, with ResNet 50 especially recognized for its consistent and high accuracy.
Examining the confusion matrix in conjunction with the training progress graph indicates that all three Convolutional Neural Network (CNN) models ResNet 50.
AlexNet, and GoogLeNet performed effectively in classifying fish species, albeit with differing degrees of accuracy and reliability.
Table 4 presents the classification outcomes for ten chosen fish species, evaluated across four anatomical regions, including whole body, head, body, and tail.
This multi-segment classification method was employed to thoroughly evaluate each modelAos robustness in recognizing both partial and obstructed views of the fish.
ResNet 50 showed a strong alignment with the dataset, accurately classifying seven out of ten fish species across all recognition segments.
Misclassifications occurred for Epinephelus areolatus.
Pampus argenteus, and Trachinotus blochii.
ResNet 50Aos strong performance across several sessions highlights its advanced architecture, enabling effective learning and generalization of intricate features in the dataset.
On the other hand.
AlexNet managed to classify just five to six species in similar conditions accurately.
Additional misclassifications were noted with Epinephelus areolatus.
Lethrinus lentjan.
Pampus argenteus.
Psettodes erumei, and Scarus ghobban.
This outcome is anticipated due to AlexNetAos relatively shallow architecture and restricted feature extraction capacities, which hinder Copyright A2025 Faculty of Fisheries and Marine Universitas Airlangga Anang et al.
/ JIPK, 17.
:608-626
its ability to discern subtle distinctions among closely related species.
GoogLeNet produced slightly better classification outcomes, accurately identifying seven to eight This improvement is linked to the modelAos unique Inception modules, which facilitate multiscale feature extraction.
However, while GoogLeNet occasionally outperformed ResNet 50 in terms of raw classification numbers, its results were less stable across repeated sessions.
Misclassifications noted included Epinephelus areolatus.
Lethrinus lentjan, and Trachinotus blochii, with only one instance occurring during the fifth training session, as indicated in Table All models exhibited substantial validation accuracy during training: ResNet 50 .
%).
GoogLeNet .
%), and AlexNet .
%), as illustrated by the training progress graphs.
These results highlight the capability of the models to generalize to new datasets, while ResNet 50 demonstrating the highest reliability and consistency as a classifier across various input perspectives and sessions.
It is crucial to understand the connection between architectural depth, training time, and classification accuracy when choosing a suitable CNN model for large-scale fish species classification endeavors.
2 Discussion 1 Discussion on the confusion table A thorough examination of the confusion matrix and training progress graphs indicates that the ResNet50 model consistently outperformed both AlexNet and GoogLeNet regarding classification accuracy, robustness, and reliability.
As highlighted in Table 5.
ResNet50 achieved nearly perfect classification across various training sessions, showing minimal misclassifications.
The modelAos capacity to accurately identify species from diverse anatomical perspectives, including the whole body, head, body, and tail, highlights its outstanding generalization capabilities, thereby making it particularly suitable for high-resolution fish Nevertheless, this remarkable performance is accompanied by considerable limitations.
ResNet50 required a significantly longer training duration of 1,346 minutes and 27 seconds, as opposed to AlexNetAos 193 minutes and GoogLeNetAos 191 minutes.
The extended duration of training can be attributed to the intricate architecture of the model and the heightened computational requirements.
An enhancement of CPU hardware may reduce this duration (Dong et , 2.
however, the employment of GPU or TPU acceleration, coupled with architectural advancements such as model pruning, dropout techniques, and batch normalization, may present more efficient and scalable solutions (Xu et al.
, 2022.
Knausgyurd et al.
, 2.
2 Discussion on the training progress graph The training progress graph reinforces these Figure 5.
shows that ResNet50 achieved over 90% validation accuracy within the first 50 iterations and reached perfect accuracy by epoch 1.
This quick convergence suggests that the model effectively captured intricate patterns and features early in training.
In contrast.
AlexNet and GoogLeNet required nearly 100 iterations to surpass the 90% validation threshold, as illustrated in Figures 5.
Although these models also attained high accuracy (AlexNet: 97.
GoogLeNet: 98.
71%), the disparity in their convergence rates emphasizes the differing capacities of each model in processing complex image In spite of their inherent strengths, all three models of Convolutional Neural Networks (CNN.
experienced difficulties with certain species, particularly Pampus argenteus, which was predominantly misclassified across all models.
This persistent misclassification suggests that the current dataset may be deficient in high-quality and diverse images pertaining to this specific species.
Factors contributing to these inaccuracies may include reflections of light, insufficient contrast, and morphological similarities to other species (Sun et al.
, 2020.
Hou et al.
, 2.
In order to address these challenges, the following enhancements are recommended: firstly, the implementation of advanced augmentation techniques, such as varying rotation angles, random cropping, contrast adjustments, and the incorporation of species-specific features, for example, body stripe coloration and tail Moreover, establishing controlled lighting during image capture and using post-processing tools to standardize brightness and reduce glare is essential.
Additionally, increasing the number of images for commonly misclassified species will promote a fairer class distribution and boost visual diversity.
ResNet50 exhibits superior classification performance in this study, corroborated by both quantitative metrics and visual convergence assessments.
Future research should concentrate on optimizing training durations and enhancing dataset quality to improve the reliability and scalability of deep learning models employed for automated fish species classification.
Conclusion In this study, three Convolutional Neural Network architectures.
ResNetAe50.
AlexNet and GoogLeNet were compared in classifying 20 local fish species.
Using a well-annotated dataset of about 20,000 underwater images captured from different an- JIPK: Scientific Journal of Fisheries and Marine JIPK Vol 17 No 3.
October 2025 | Deep Learning Models Performance on Marine Fish Species Classification atomical perspectives, all proposed models achieved high classification accuracy.
Among the three proposed models.
ResNetAe50 had the highest classification accuracy.
This is likely due to the residual skip connections that help mitigate degradation in deeper The minor misclassifications that did occur like in Pampus argenteus were likely a result of overlapping visual features and/or inconsistent image Most importantly, current work shows that a well-labeled dataset is essential for a model to identify more subtle features and symmetries, rather than depending solely on its architecture.
Without enough labelled data, even state-of-the-art architectures may struggle to achieve good classification performance in heterogeneous underwater environments.
Results demonstrated that deep learning is a viable and robust method to classify underwater species if a sufficient labelled dataset is available.
Going forward, results identified several major research gaps: .
broader taxonomic coverage, i.
, more species and environments, .
optimization for real-time deployment .
, transfer learning, lightweight architectur.
, and .
seamless integration of classification models into fisheries monitoring systems.
Resolutions to these gaps can help improve adaptability to varying ecological contexts.
Looking ahead, our results highlight a few exciting areas for further research: .
expanding our taxonomic coverage by including more species and environments.
improving models for real-time application using techniques like transfer learning and lightweight architectures.
integrating classification models into fisheries monitoring systems to make them even more effective.
Collectively, these directions can improve adaptability across various ecological settings and resource-limited environments.
connecting deep learning with fisheries science, this study lays the foundation for better ecosystem monitoring, sustainable fisheries management and marine biodiversity conservation.
Future efforts should aim to improve the modelsAo capacity to generalize, transfer knowledge and perform inference directly on devices in order to better address the complex requirements of marine resource management.
Acknowledgement We thank Universiti Malaysia Terengganu for providing funding support for this project (UMT/ TAPE-RG/2023/55.
AuthorsAoContributions All authors have played crucial roles in the creation of the final manuscript.
Their specific contributions are detailed as follows: Nur Muhammad Afiq Anang gathered the data, executed data augmentation, carried out data processing, managed the training process, and composed the manuscript.
Ezmahamrul Afreen Awalludin refined the main conceptual ideas of the article, organized its structure, and offered comments and suggestions for improvement.
In addition, all authors collaborated to discuss the results and enhance the final manuscript together.
Conflict of Interest The authors hereby declare that no conflicts of interest are associated with this publicationAos writing.
Declaration of Artificial Intelligence (AI) The author confirms that this manuscript is entirely composed and developed independently.
However, the author has used Grammarly software to correct the grammar, sentence structure, and use of appropriate words in the manuscript.
All content presented here is assumed to be their own creativity and dedication, demonstrating its originality and integrity.
Funding Information This research was supported by UMT (UMT/ TAPE-RG/2023/55.
References