Spatial attention–driven hybrid Vision Transformers for automated classification of crown-of-thorns starfish in coral reef ecosystems

1 Introduction

Coral reefs are well known for their vibrant color. Moreover, there are often seen as a valuable species on the earth. Coral reefs are beautiful, colorful ecosystems compared to the rainforests of the sea because of the marvelous biodiversity they support. Even though they occupy less than one percent of the ocean floor, reefs provide shelter, food, and a place for about a quarter of all marine species. These reefs mostly grow from hard calcium carbonate and build complex three-dimensional structures that support a very sensitive network of fish, algae, shellfish, and small, microscopic animals. Coral reefs are very important as the natural first line of defense, absorbing up to ninety percent of wave energy before it actually gets to shore and hence preventing coastlines from expensive damage through erosion and flooding. In return, reefs have a role in balancing Earth’s carbon and nitrogen cycles by the absorption of CO₂ while maintaining the chemical equilibrium of oceans. Apart of their value in terms of ecology, coral reefs are huge contributors to the economy. Tourists visit these places having coral reefs for surfing purpose and to view the colorful coral reefs. Some locals mainly depend on tourism income. But this great resource is significantly under the threat. Global Warning results in corals bleaching, meaning the corals lose its color. Pollution and ocean acidification also pose harm them. The big natural hazard to the coral reefs is the Crowns of Thorns Starfish. This starfish which has large spines feeds on the coral reefs. In a population explosion of more than a thousand per hectare, however, they can devastate reefs. These starfish—up to forty centimeters wide with twelve to twenty-three venomous arms—consume several square meters of coral per year. During outbreaks, Crown of Thorns Starfish can rapidly consume whole reefs. They feed by turning their stomachs inside out and applying enzymes that break down the coral tissue, which results in white, naked skeletons. This damage to coral homes for many fish and small sea life leads the reef ecosystem to decline. Recovery may take years or decades depending on the type of coral, water quality, and regrowth success. Repeated outbreaks might permanently shift healthy, colorful coral reefs into areas mainly ruled by algae, which disrupts both biodiversity and food chains. Overfishing of natural predators such as the Giant Triton snail, plus nutrient pollution and warming seas, is a way for starfish to overpopulate by improving their survival at a young stage. This urges us for early detection as well as regular monitoring that will help in coral reefs restoration. Manual monitoring techniques like Manta Tow and diver surveys and the underwater photography can result in human errors. This makes it necessary for an automated intelligent monitoring system. Deep Learning and computer vision have greatly enhanced the detection and classification of underwater images accuracy. Although Convolutional Neural Networks (CNNs) are normally used for feature extraction, they come with a high requirement for well-labeled data as well as significant computational resources. Also, long-range spatial relationships are not easily captured by CNNs, while details are heavily masked due to such underwater conditions as turbidity, varying light conditions, and sediment. Preference goes to Transformer-based models because they keep up great fine details—in addition to maintaining the overall scene structure—with attention to important regions of the image. They work very well in challenging underwater low-visibility and lighting conditions. This study proposes a deep hybrid learning methodology for automatic identification and classification of underwater organisms with specific orientation toward the Crown of Thorns Starfish based on three transformer architectures—Swin Transformer, FastViT, and CrossViT. Whereas earlier models only recognized the presence of starfish, our model classifies images into five categories: Acanthaster planci (COTS), Echinoidea (Sea Urchin), Crinoidea (Feather Star), Pycnopodia helianthoides (Sunflower Starfish), and Scleractinia (Coral Reefs). Species with very similar texture and symmetrical forms mostly bewilder not-so-intelligent algorithms; therefore, category selection has been done accordingly. The CNN-based models achieved accuracy ranging from 91.80% to 95.90%, while the transformer-based models achieved accuracy between 97.54% and 97.81%. The proposed Spatial-Attention Swin + FastViT + CrossViT framework achieved the highest accuracy of 99.18% on the marine species classification task. It mixes up big and small picture looks for spotting things right in many underwater scenes. Because it works well and stays strong, this makes it a hopeful choice for wide reef checks to save the world’s valuable coral life spots. The remaining section of this manuscript are as follows: Section 2 describes the existing works, Section 3 describes the materials and methodology, and Section 4 illustrates the experimental results and discussion, Section 5 illustrates Ablation Study and Section 6 describes Conclusion and Future Enhancement.

2 Related works

Recently, multiple scientists have shifted their focus towards the development of intelligent systems for the identification of the Crown-of-Thorns Starfish (COTS), a major species that destroys coral reefs. Coral reefs serve as the foundation of marine biodiversity, which has led researchers and engineers to apply deep learning and computer vision methods to address the detection capabilities of COTS in complex underwater environments. Although there have been several prior studies, one of the more notable examples was by Truong (2022), who used a YOLOv5 model to identify COTS from underwater images, which was able to successfully identify COTS in near real-time, even in turbid waters and poor lighting. Similar to that of Truong, Sheth and Prajapati (2022) used CNNs to successfully identify COTS from other marine species, demonstrating how trained neural network models can identify subtle differences between similarly aquatic organisms. In addition to that, Nguyen (2022) also improved the YOLOv5 model to use TensorFlow Lite for smaller embedded systems to be able to inform on-site detection with low-power device on-site. The advancement of Heenaye-Mamode Khan et al. (2023) took it a step further by embedding an attention mechanism into the CNNs to help the network focus on the distinctive spiny elements which are the identifying features of COTS. Relano et al. (2024) subsequently introduced Vision Transformers (ViTs) for underwater classification efforts; they demonstrated that models based on self-attention can consider both large-scale spatial features as well as fine surface characteristics with greater accuracy than CNNs. In a later effort, Dubey and Nirmal (2024) also improved recognition accuracy to intelligently detect COTS, improved by analyzing multi-scale textures that were more indicative of their complex structures inherent with marine starfish. Li et al (Li et al., 2023). proposed a method that combined sensor-based environmental features with image-based inputs to predict potential outbreaks, while Tarafdar et al. (2024) proposed a similar machine learning model with ecological data for sustainable reef management. The extensive ecological observations completed by Pratchett et al. (2017) provided long-term insights into outbreaks of starfish which helped to influence model-based AI control strategies in their studies. Dumas et al. (2020) recommended using citizen-scientist-collected reef imagery as an efficient mechanism for training deep-learning systems. Robotics also began to form a key component in reef protection efforts beginning with Dayoub et al. (2015), who tested submerged robots with a visual detection system to autonomously monitor and record the presence of COTS over reef zones connected AI detection results with coral health assessments, demonstrating how data-driven approaches could be beneficial for monitoring the decline of those ecosystems. Pooloo et al. (2021) investigated the use of artificial intelligence for monitoring causes associated with coral reef degradation. Cherian et al. (2022) stressed the importance of image enhancement prior to learning model use, as underwater images may misrepresent color and clarity. Dakhil and Khayeat (2022) discussed some general challenges with underwater detection and presented several limitations to underwater detection in general such as inconsistent light and inadequate labeled training datasets. Addressing these obstacles, Feng and Jin (2024) proposed CEH-YOLO, a modified object detection model that worked well under turbid conditions, while Liu et al. (Liu et al., 2019) presented a YOLOv3-GAN hybrid framework that could focus more on the marine features rather than the high variability of noise in the background. Li et al. (2022) designed an edge-based AI reef monitoring device that is able to process underwater data in real time, without server intervention. Zhang et al. (2022) addressed the challenge of mapping coral shapes and surfaces by using deep learning and semantic segmentation, providing a more detailed ontology to coral formations. Saleh et al. (2022) showed how CNNs could be developed for classifying fish and advanced early attempts to incorporate image analysis to aquatic animals. Mathur and Goel (2021) used FishResNet, a computer vision model based on CNNs. Rathi et al. (2017) applied convolutional neural networks and deep learning for underwater fish species classification, while Li et al (Li et al., 2023). reviewed deep-learning approaches for the visual recognition and detection of aquatic animals. Li and Du (2022) explored the application of deep learning in marine biology. They indicated that deep learning could be useful in aquaculture and habitat assessment. Zhao et al. (2024) reviewed deep-learning-based stereo vision techniques for phenotype feature and behavioral analysis of fish in aquaculture. Ziaei et al. (2025) investigated image-based individual fish identification as a substitute for invasive identification methods. Jalal et al. (2025) introduced DeepFins, an approach for fish detection by capturing dynamics in underwater videos.

Liu et al (Liu et al., 2025). suggested the YOLO-LDFE (Liu et al., 2025) approach, which performed well with respect to speed and accuracy. Mehrunnisa et al. (2025) modified ResNet-18 to obtain better performance in darker conditions when images were taken underwater. Cao et al. (2025) created DNnet which corrected underwater images by correcting for the coloration of images to obtain clarity for underwater images. Dürrani et al (Dürrani et al., 2025). created a classification approach with MobileNet and MLP-Mixer which was able to classify marine organisms with a simple and small device. Muntaner-Gonzalez et al. (2025) developed a segmentation system for fish prior to carrying out classification. Lawal et al. (2025) presented the MADNet approach which provided faster classification for marine organisms. Ajay et al. (2025) proposed FishFusionNet (Ajay et al., 2025), which took advantage of attention with SqueezeNet. Zhao and Dong (2025) applied a graph-based attention approach that could detect fish regardless of busy backgrounds. Pratt and Mandrak (2025) captured marine data using batch images without causing harm to fish in marine environments. Ogidi et al. (2025) discussed the role of computer technology in identifying marine organisms. Qian et al. (2026) proposed HiFi-Det, which used collaborative multi-scale frequency-domain feature optimization for Crown-of-Thorns Starfish detection in complex underwater environments. Bhat et al. (2025) combined two or more models and improved [they] detection rate of marine organisms in ecosystems with little visibility. Zhang et al. (2025) adapted VGG-16 using fish images.

Myrvold et al. (2025) used underwater videos and computer vision to monitor fish assemblages in seasonal off-channel habitats. Jyothimurugan et al. (2025) used a deep convolutional generative adversarial network with YOLOv6 and proposed an embedded artificial intelligence framework for detecting COTS in underwater imagery. Lopez-Vazquez et al. (2023) applied deep-learning-based image enhancement and animal species classification to deep-sea imagery. Jiao et al. (2025) developed the DS-YOLOv8 method for underwater biological detection. Gollapalli et al. (2024) designed an application for monitoring Crown of Thorns (COTS) in real reefs. Gorman et al. (2025) examined biologically based technological developments for the management of Crown-of-Thorns sea stars. Talele and Patil (2022) integrated visual attention method and detection of starfish on video filmed with movement. Krishnan et al. (2022) combined underwater object detection and tracking using a hybrid deep convolutional neural network. Roy and Talukder (2024) created an efficient model for many marine organisms in various light and textures. Alkhalifa et al. (2025) presented a hybrid deep-learning framework incorporating the Starfish Optimization Algorithm for secure intelligent edge computing. Yang et al. (2019) were among the first to use CNNs for marine animals. Hu et al. (2020) developed the “Starfish” method for compressing many underwater images. Liu et al (Liu et al., 2025). proposed an enhanced Starfish Optimization Algorithm using a joint strategy and applied it to ultra-wideband indoor positioning. Elawady (2015) developed one of the earliest CNN models based on coral images and is a pioneer in this field.

3 Materials and methods

3.1 Dataset description

In our study, we have used a total of 2,438 underwater photographs of marine species, as shown in Table 1. The dataset contains five classes namely the Crowns of Thorns Starfish (COTS), Sunflower Starfish, Feather Starfish (Crinoids), Sea Urchin, and Coral Reefs. This dataset was obtained from Kaggle. To ensure reliable model training and unbiased evaluation, the dataset was divided into 70% training (1,706 images), 15% validation (366 images), and 15% testing (366 images) using a stratified sampling strategy, which preserved the class distribution across all subsets. The reason why we chose these five classes is because they look quite similar to COTS, which makes it tough to classify them in underwater and dark environments. The Feather Stars have long, soft arms that spread like a star, and Sunflower Starfish have many rough arms that can also look like COTS. Even Coral Reefs, though they do not move, have rough surfaces and mixed colors that can resemble hidden starfish. Using these similar-looking marine organisms helped the model learn subtle visual differences and improved their accuracy in correctly identifying the Crown-of-Thorns Starfish. Furthermore, the inclusion of visually challenging and morphologically similar classes enhances the robustness and generalization capability of the proposed model for real-world underwater monitoring applications. The sample images from the marine organism dataset are shown in Figure 1.

Table 1. Summary of the underwater marine species dataset showing the class-wise image distribution and the corresponding training, validation, and testing splits.

3.2 Data augmentation

The data used for the study which includes around 2,438 underwater images was broken down into five biological categories: Crown of Thorns Starfish (Acanthaster planci), Sea Urchin (Echinoidea), Feather Star (Crinoidea), Sunflower Starfish (Pycnopodia helianthoides), and Coral Reef (Scleractinia). Some of the image categories have fewer images than others, warranting additional image generation to obtain a more balanced and representative dataset. Various transformations were applied to the images to generate a more diverse representation of the image properties that would typically be found during filtration. The modified images were subjected to no less than 5 individual transformations; including horizontal and vertical reflection to mimic physical viewpoint, gentle rotations of no more than 0.2 radians to mimic potential alignment in camera location, and a minor degree of zoom of no more than 0.2 to imitate distance from the organism occurred relative to the lens of the camera. The sample data-augmented underwater images used to enhance model robustness are shown in Figure 2. Small translations of 0.2 - both horizontally and vertically - repositioned the organisms within the image while small adjustments of 0.2 contrast altered lighting or color, replicating the effects of depth, turbidity or water movement. The dataset is first divided into training, validation, and testing subsets using stratified sampling. Data augmentation is employed only for training subset after dataset partitioning. No original image or augmented version appears in more than one subset, thereby preventing data leakage during model evaluation.

3.3 Data splitting

The dataset contained around 2,438 images which were later divided into three groups to ensure fair evaluation. Around 70% was used for training, 15% for validation, and the remaining 15% for testing. This was done so that the model learned effectively while being tested on unseen data. The images that were trained helped the model identify patterns and adjust its internal parameters accordingly, while the validation images helped to fine-tune and avoid overfitting. The test images were not used while training to measure final performance objectively. Each split had an equal mix of all five species under various lighting and orientation conditions. This ensured that the proposed model was able to perform well on real-world underwater images. This method helped to build a foundation for testing. The general system architecture for the marine species classification framework is shown in Figure 3.

3.4 Model description

3.4.1 VGG19

VGG19 is one of the deep neural network models for image classification. This model comprises convolutional and fully connected layers that sum up to a total of nineteen layers. Among its architectural features, what makes it outstanding is the use of small 3×3 filters which enables the model to capture very fine visual details and keeps the general architecture clean and consistent. After some convolutional layers, there are pooling layers that reduce dimensional data so focused on important features. As noticed in the depth of VGG19, only simple edges and outlines can be found at shallower depths, but more detailed texture and shape information is discovered at increased depths. The design has been quite appreciated due to its simplicity and ease of stable training while attaining high performance in meaningful feature extraction from images.

The Convolution and ReLU operation is mathematically is defined in Equations 1, 2.

(1) Cost(convolution)≈H·W·K2·Cin·Cout

(2) ReLU(x)=max(0,x)

3.4.2 MobileNetV3

MobileNetV3 is lightweight and is designed for speed. This is suitable for mobile device deployment using limited computational resources. This uses Dept wise Separable Convolutions and Squeeze-and-Excitation Blocks which help improve enhancement of the feature. This also uses Hard-Swish which results in both accurate and smooth training.

The computational cost comparison between a depthwise separable convolution and standard convolution is given in Equations 3, 4.

(3) Cost (standard convolution)≈H·W·K2·C_in·C_out

(4) Cost(depthwise separable convolution)≈H·W·(K2·C_in+C_in·C_out)

where H, W = spatial height and width, K = kernel size, C_in, C_out = input and output channels.

The hard-swish activation is defined in Equations 5, 6.

(5) h−swish(x)=x·ReLU6(x+3)/6

where.

(6) ReLU6(x)=min(max(0,x),6)

3.4.3 DenseNet

DenseNet is a type of deep learning architecture. It’s known for connecting each layer to every other layer in a feed-forward fashion. DenseNet connects each layer to all preceding layers within the same block, thereby allowing information and gradients to flow unimpeded throughout the network- this helps prevent loss of data during training. Instead of relearning identical features, each layer makes use of features extracted by all its preceding layers, making it more efficient and less redundant. Dense blocks are separated by transition layers that assist in keeping the model compact and stable. This connectivity enables DenseNet to attain high accuracy with even fewer parameters plus fast training and it is represented in Equation 7.

(7) xl=Hl([x0,x1,…,xl−1])

where [x0, x1,…, xl − 1] represents the concatenation of feature maps from previous layers, Hl is a composite function of batch normalization, ReLU, and convolution.

3.4.4 ResNet

ResNet (Residual Network) altered deep learning by adding skip connections, allowing the flow of information not to be blocked by some layers. This novel technique addressed the vanishing gradient problem that made training deep networks unstable. It does not learn the whole transformation; each layer learns only the difference between input and output. Therefore, training becomes faster and more stable. Due to this easy but strong concept, networks with a great depth of layers can be trained efficiently and accurately. ResNet is a baseline model for computer vision. The operation within a residual block can be mathematically represented in Equation 8.

(8) H(x)=F(x)+x

where x is the input to the block, F(x) represents the residual function learned by the stacked convolutional layers, H(x) is the final output of the block.

3.4.5 CrossViT

CrossViT gives two side-by-side ways for picture pieces of different sizes. In one-way, small pieces show fine details while in the other way, bigger pieces look at general context. These two ways share information by cross-attention connecting detailed textures to the general structure as stated in Equation 9. With a dual-scale setup, this model can catch well both local and global patterns giving good exactness at fair computing needs.

(9) CrossAttention(Qs,KL,VL)=Softmax((QsKLT)/√dk)VL

where Qs , KL , and VL denote the query, key, and value matrices of small and large patch streams, dk is the key dimension.

3.4.6 FastViT

FastViT is a model that mixes ideas from CNN and Transformers. Normal Transformers are slow, but FastViT removes the heavy and slow parts while keeping the important details from the image. It joins image pieces step by step, so the computer has less work to do. This makes FastViT fast and still accurate, even on small or weak devices. So, it is good for real-time image work where we need both speed and good results.

3.4.7 Swin Transformer

Swin Transformer looks at an image in small, fixed windows, like looking through small boxes. First it only focuses on what is inside each box, so it saves computation. Then in the next layer, the boxes move, so it can also see nearby parts and connect them. It also combines patches so it can learn both small details and big shapes. This mix of close and wide view makes Swin Transformer very good for detailed image understanding. The cost for the Swin Transformer is calculated using Equation 10.

(10) Cost(global attention)≈(H·W)2·C

where H, W = spatial height and width of the feature map, M = window size, C = feature dimension.

3.4.8 Proposed Spatial-Attention Swin-FastViT-CrossViT model

The proposed Spatial-Attention Swin-FastViT-CrossViT model combines three transformer backbones to capture complementary visual features for marine species classification. First, a separate spatial attention module is applied to each branch. Channel-wise average and maximum maps are combined using a 7×7convolution and sigmoid activation to generate the attention map as stated in Equations 11, 12.

(11) A(x)=σ(Conv7×7[Avg;(x)Max(x)])

and the attended image is obtained as.

(12) xatt=x⊙(0.5+A(x))

The attended images are processed independently by Swin-Tiny, FastViT-T8, and CrossViT. Their extracted features are standardized and projected into a common 256-dimensional feature space, producing z_S, z_F, and z_C. These features are combined using adaptive gated feature fusion, where the contribution of each backbone is learned in Equations 13, 14.

(13) [gS,gF,gC]=Softmax(G([zS;zF;zC]))

and the final fused representation is.

(14) zfused=gSzS+gFzF+gCzC

Finally, the 256-dimensional fused feature is passed through Layer Normalization, a 128-neuron fully connected layer with GELU activation, Dropout (0.30), and a five-class Softmax output layer as stated in Equation 15.

(15) y^=Softmax[W2Dropout0.30(GELU(W1LN(zfused)+b1))+b2]

This process combines spatially important image regions with complementary features from the three transformer backbones to produce the final marine species classification. Algorithm 1 represents the pseudocode for the proposed algorithm.

Algorithm 1

Figure 4 represents the workflow of the proposed hybrid model for marine species classification. The experimental design for this proposed research work is shown in Table 2. The proposed framework applies spatial attention independently to the Swin-Tiny, FastViT-T8, and CrossViT branches to obtain discriminative feature representations. The extracted features are standardized, projected into a common 256-dimensional space, and combined using adaptive gated feature fusion, followed by LayerNorm, a GELU-activated fully connected layer, dropout, and Softmax for final five-class marine species classification.

Table 2. Experimental settings.

4 Results and discussion

The effectiveness of the developed hybrid deep learning model was evaluated using widely adopted classification performance metrics, including accuracy, precision, recall, and F1-score. Table 3 presents the test-set classification performance comparison of the evaluated models on 366 test images using accuracy, precision, recall, and F1-score. To provide a more comprehensive assessment of the transformer-based and hybrid architectures, additional evaluation metrics including macro precision, macro recall, macro F1-score, ROC-AUC, Matthews Correlation Coefficient (MCC), and Cohen’s kappa (κ) are reported in Table 4. These complementary metrics provide a more detailed analysis of the classification performance, robustness, and agreement of the proposed framework across multiple classes.

Table 3. Test-set classification performance comparison of the evaluated models on 366 test images using accuracy, precision, recall, and F1-score.

Bold values represent that the proposed algorithm performs better.

Table 4. Extended test-set performance comparison of the evaluated models on 366 images using macro precision, macro recall, macro F1-score, ROC-AUC, MCC, and Cohen’s κ.

Bold values represent that the proposed algorithm performs better.

As shown in Table 3, the conventional CNN-based models, including VGG19, MobileNetV3-Small, DenseNet121, and ResNet50, achieved lower classification accuracy ranging from 91.80% to 95.90% compared with the transformer-based architectures. Among the individual Vision Transformers, FastViT-T8 achieved the highest accuracy of 97.81%, followed by Swin-Tiny (97.54%) and CrossViT (97.54%). The hybrid transformer models further improved the classification performance by combining complementary feature representations, with Swin-Tiny + CrossViT achieving 98.91% accuracy. The proposed Spatial-Attention Swin + FastViT + CrossViT framework achieved the best overall performance with 99.18% accuracy, 99.18% precision, 99.18% recall, and 99.18% F1-score, demonstrating the effectiveness of spatial-attention-driven multi-transformer feature fusion in improving marine species classification under challenging underwater conditions.

The detailed performance analysis presented in Table 4. The proposed model achieved the highest Macro Precision of 99.36%, Macro Recall of 99.23%, and Macro F1-score of 99.30%, indicating balanced classification performance across all five marine species. Furthermore, it obtained a ROC-AUC of 99.99%, MCC of 0.9894, and Cohen’s κ of 0.9894, outperforming all individual transformer and hybrid models. These results demonstrate that the spatial-attention-based feature fusion strategy effectively combines the complementary strengths of Swin Transformer, FastViT, and CrossViT, resulting in highly discriminative feature representations, improved classification robustness, and excellent agreement between predicted and ground-truth labels. Overall, the proposed framework provides a reliable and efficient solution for automated underwater marine species classification and Crown-of-Thorns Starfish detection.

Table 5 compares the number of correct predictions and classification errors obtained by the evaluated models on the test set. The CNN models produced comparatively more errors, ranging from 15 to 30, while the individual transformer and hybrid models reduced the errors to 4–9. The proposed Spatial-Attention Swin + FastViT + CrossViT model achieved the best result, with 363 correct predictions and only 3 misclassified images, demonstrating its improved classification performance over all other evaluated models.

Table 5. Test-set comparison of correct predictions and classification errors of the evaluated models across 366 images.

Bold values represent that the proposed algorithm performs better.

Figures 5–12 present the training behavior, classification performance, and feature representation of the evaluated models. Figure 5 illustrates the learning curves of the proposed framework in terms of accuracy, loss, macro precision, macro recall, macro F1-score, and MCC, providing an overview of its training progression. Figures 6–10 present the confusion matrices of the CNN, individual transformer, and hybrid transformer models on the test images, showing the distribution of correct and incorrect predictions across the five marine species. Figure 11 presents the confusion matrix of the proposed Spatial-Attention Swin + FastViT + CrossViT model, which correctly classified 363 images with 3 misclassifications. Figure 12 presents the t-SNE visualization of the final fused feature representations of the proposed framework, showing the distribution and clustering of the learned features across the five marine species classes.

Table 6 presents the five-fold cross-validation results of the proposed spatial-Attention Swin + FastViT + CrossViT framework. The proposed model demonstrates consistently high performance across all five folds, with only minor variations in the evaluation metrics. The consistently high accuracy, ROC-AUC, macro and weighted metrics, MCC, and Cohen’s κ indicate that the proposed framework learns robust and discriminative feature representations while maintaining excellent generalization across different data partitions. These results confirm that the adaptive gated feature fusion strategy provides stable, reliable, and consistent classification performance without significant performance degradation across folds.

Table 6. Performance metrics obtained across the five-fold cross-validation experiments for the proposed model.

Table 7 summarizes the average performance of the proposed Spatial-Attention Swin + FastViT + CrossViT framework across the five-fold cross-validation experiments. The model achieved a mean accuracy of 98.31 ± 0.87% and a ROC-AUC of 99.90 ± 0.09%. The low standard deviations across all evaluation metrics indicate consistent performance over different folds. Similarly, the macro and weighted precision, recall, and F1-score values show only small variations, while the MCC (0.9782 ± 0.0112) and Cohen’s κ (0.9781 ± 0.0113) indicate a high level of agreement between the predicted and true class labels across the five validation folds.

Table 7. Average performance (mean ± standard deviation) of the proposed model across five-fold cross-validation.

Table 8 presents the class-wise performance of the proposed model across the five-fold cross-validation experiments. The Coral Reefs class, with the highest average support (137.0 samples), achieved an F1-score of 98.77 ± 0.79%. Sunflower Starfish achieved the highest F1-score (99.03 ± 1.04%), while COTS, Crinoids, and Sea Urchin achieved F1-scores of 97.84 ± 1.55%, 97.83 ± 1.84%, and 97.74 ± 2.23%, respectively. The relatively low standard deviations across all classes indicate consistent class-wise performance over different validation folds. The training and validation accuracy and loss curves of the proposed framework across the five-fold cross-validation experiments (Folds 1–5) is illustrated in Figure 13.

Table 8. Per-class performance analysis (mean ± standard deviation) of the proposed model across five-fold.cross-validation.

Figure 14 illustrates the occlusion sensitivity analysis of the proposed model and predicted output of the proposed model. Occlusion sensitivity is an explainable AI (XAI) technique used to identify the image regions that have the greatest influence on the model’s prediction. By systematically masking different parts of the input image and observing the change in prediction confidence, it highlights the most discriminative features used for classification. The heatmaps show that the proposed model primarily focuses on the characteristic regions of each marine species, indicating that the predictions are based on relevant visual features rather than background information.

The sample predictions and corresponding confidence scores for selected test images are shown in Figure 15. Crown-of-Thorns Starfish – 99.98%, Coral Reef – 99.84%, Sea Urchin – 99.73%, Sunflower Starfish – 97.20%, and Feather Star – 74.33%.

5 Ablation study of the proposed model

Table 9 presents the ablation study of the individual transformer backbones and their fusion combinations. Among the standalone models, FastViT-T8 achieved the highest performance (97.81% accuracy). The two-model fusion Swin-Tiny + CrossViT further improved the results to 98.91% accuracy. The proposed Spatial-Attention Swin + FastViT + CrossViT framework achieved the best overall performance with 99.18% accuracy, 99.30% Macro F1-score, and 0.9894 MCC, indicating that combining complementary features from all three backbones improves classification performance over the individual and dual-model variants.

Table 9. Ablation study of the individual transformer backbones and their fusion combinations.

The comparative study of the existing and the proposed works is presented in Table 10. In the existing work by Heenaye-Mamode Khan et al. (2023), the authors focused on binary classification of starfish, specifically distinguishing between COTS and non-COTS images. They implemented transfer learning using pre-trained CNN architectures such as VGG19 and MobileNetV2. The baseline VGG19 model achieved an accuracy of 81.3%, which was then optimized using the Adam optimizer to reach 87.1% accuracy. Furthermore, an enhanced VGG19 with a Convolutional Block Attention Module (CBAM) was proposed, improving classification performance to 92.6% accuracy and 95% mean Average Precision (mAP). Their approach successfully demonstrated that adding attention to CNN increases feature interpretability and improves COTS vs non-COTS recognition. Table 9 further illustrates the progressive improvement in performance. The multiresolution Colored Wavelet-guided GLCP with XGBoost proposed by Dubey and Nirmal (2024) achieved an accuracy of 95.00%, precision of 0.98, recall of 0.90, and F1-score of 0.94, demonstrating the effectiveness of multiresolution handcrafted feature extraction for binary COTS classification. However, the system was limited to binary classification, relied only on CNN-based models, and did not include multi-class marine species such as urchins, sunflower starfish, crinoids, or coral reefs. Moreover, the reliance on a single CNN architecture limits the model’s ability to effectively capture complementary local and global feature representations required for distinguishing visually similar marine organisms. In contrast, the proposed hybrid model in this research extends the problem from binary COTS identification to five-class classification (COTS, crinoids, sunflower starfish, urchins, and coral reefs) in real underwater conditions. These additional classes were specifically chosen because they visually resemble COTS in terms of shape, structure, and color patterns, which often leads to misclassification in real-world reef environments if only binary classes are used. Instead of using a single CNN network, a transformer-based hybrid fusion of Spatial-Attention Swin + FastViT + CrossViT is implemented. Each model independently extracts global pooled features, which are concatenated and passed to a fully connected classifier. The proposed framework achieved an overall accuracy, precision, recall, and F1-score of 99.18%, outperforming the existing classification approaches while effectively handling the more challenging five-class marine species classification task.

Table 10. Comparison of proposed model with existing COTS classification approaches.

Bold values represent that the proposed algorithm performs better.

6 Conclusion and future enhancement

The present study introduced the Spatial-Attention Swin + FastViT + CrossViT hybrid framework for efficient classification of marine species with emphasis on detection of Crown-of-thorns starfish in aquatic contexts. The efficiency of the suggested framework can be exemplified by its overall classification accuracy of 99.18%, which is remarkable compared to CNN-based models and other transformer approaches examined. Also, Explainable AI was implemented in this study by means of applying occlusion sensitivity analysis so that image areas responsible for classification process could be visualized, and it was found that the proposed approach concentrates on biologically meaningful properties of the images instead of using irrelevant background information. Therefore, it can be concluded that the proposed approach can be efficiently adopted in automated underwater coral reefs’ monitoring and early detection of Crown-of-thorns starfish outbreaks. Further research will focus on enlarging the dataset with respect to the diversity of images utilized in this study considering involvement of more species as well as making it usable for robotics and edge-AI applications.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Tim Wilhelm Nattkemper, Bielefeld University, Germany

Reviewed by: Younes Ouassine, Institut de Recherche pour le Developpement au Maroc, Morocco

Lotfi Ezzeddini, Faculty of Economics and Management, Tunisia