Deep learning-based classification of species in central-southern fisheries in Chile

Abstract

This study introduces a novel deep learning methodology for identifying fish species in central-southern Chile's pelagic and demersal fisheries. Using a dataset of 8,118 high-resolution images encompassing 18 species, two Convolutional Neural Networks (CNNs) were developed: a custom-designed CNN, which achieved an overall accuracy of 86% (95% CI: [0.8355; 0.8826]), and an adapted VGG16 model, which reached 95% (95% CI: [0.9355; 0.9651]) when tested on the same set of 811 images. While both models perform strongly, challenges persist for specific species, particularly Brama australis and Strangomera bentincki, with 33 and 53% classification rates in the VGG16 model, highlighting opportunities for dataset enrichment and algorithmic refinements. Additionally, Gradient-weighted Class Activation Mapping (Grad-CAM) was employed to visually interpret the decision-making process of the CNN, providing insight into the regions of the image most relevant to classification. Developed using the Keras API and TensorFlow framework within the R programming environment, our approach underscores the importance of advanced computational tools in enhancing species classification. The results lay the groundwork for future expansions into comprehensive frameworks utilizing computer vision to recognize fish species on board, quantify catches, and detect discards and bycatch. These advancements could significantly benefit Fisheries Observer programs, enhancing enforcement and aiding sustainable fisheries management. Ultimately, this work promotes efficiency and efficacy in monitoring, fostering a sustainable future for marine biodiversity in Chile and potentially other regions and wider marine ecosystems.

Headnote

ABSTRACT. This study introduces a novel deep learning methodology for identifying fish species in central-southern Chile's pelagic and demersal fisheries. Using a dataset of 8,118 high-resolution images encompassing 18 species, two Convolutional Neural Networks (CNNs) were developed: a custom-designed CNN, which achieved an overall accuracy of 86% (95% CI: [0.8355; 0.8826]), and an adapted VGG16 model, which reached 95% (95% CI: [0.9355; 0.9651]) when tested on the same set of 811 images. While both models perform strongly, challenges persist for specific species, particularly Brama australis and Strangomera bentincki, with 33 and 53% classification rates in the VGG16 model, highlighting opportunities for dataset enrichment and algorithmic refinements. Additionally, Gradient-weighted Class Activation Mapping (Grad-CAM) was employed to visually interpret the decision-making process of the CNN, providing insight into the regions of the image most relevant to classification. Developed using the Keras API and TensorFlow framework within the R programming environment, our approach underscores the importance of advanced computational tools in enhancing species classification. The results lay the groundwork for future expansions into comprehensive frameworks utilizing computer vision to recognize fish species on board, quantify catches, and detect discards and bycatch. These advancements could significantly benefit Fisheries Observer programs, enhancing enforcement and aiding sustainable fisheries management. Ultimately, this work promotes efficiency and efficacy in monitoring, fostering a sustainable future for marine biodiversity in Chile and potentially other regions and wider marine ecosystems.

Keywords: species recognition; deep learning; bycatch; fish management; Convolutional Neural Networks

INTRODUCTION

Bycatch-the unintentional catch of non-target species within fishing operations-remains a critical issue in marine conservation and sustainability due to its capacity to degrade marine ecosystems (Kelleher 2005, Soykan et al. 2008, Davies et al. 2009, Komoroske & Lewison 2015). Discards, a subset of bycatch, refer to the fraction of the overall catch thrown back into the sea (Kelleher 2005). Because discarding can be a wasteful practice and conflict with sustainable fishing objectives, the FAO Code of Conduct for Responsible Fisheries (FAO 1995) has articulated guiding principles and international standards to address its impacts. These include: 1) considering discard impacts when applying precautionary approaches to fisheries management; 11) systematically collecting and reporting discard-related information; iii) collecting high-quality ecosystem and fisheries data to assess discard status; iv) promoting technological and operational methods to minimize discards; and v) implementing observer and inspection programs to foster compliance with management measures.

In Chile, fishing activities contribute nearly 0.7% to the national gross domestic product, with total landings averaging around 2 million tons over the past five years (SUBPESCA 2022). Of this, 60% originate in the central-south region. Pelagic species-such as anchovy (Engraulis ringens), common sardine (Strangomera bentincki), and jack mackerel (7rachurus murphyi) account for roughly 56% of these landings, while demersal species-like common hake (Merluccius gayi) and hoki (Macruronus magellanicus) contribute about 4%. Under Chile's General Law of Fishing and Aquaculture, discards must be included when determining annual global catch quotas (Paragraph 1 Bis, Law N°20.625), a critical measure to ensure sustainable exploitation of marine resources. Although global bycatch rates have been estimated to reach 40.4% (Davies et al. 2009), discard rates in some Chilean fisheries, such as the southern hake (Merluccius australis), were estimated to be around 19% during 2000-2003. To address such challenges, Chile began incorporating bycatch and discard observations into its fisheries monitoring programs in 2014, starting with pelagic and extending to demersal fisheries. By 2017, bycatch and discard reduction plans had been formally integrated into observer and monitoring programs for major fisheries, implemented by the Instituto de Fomento Pesquero (IFOP), involving administrative and conservation measures, the adoption of technologies aimed at reducing discards, robust monitoring programs, the evaluation of effectiveness of these measures, training and awareness initiatives, and a code of best practices (Roman et al. 2022, Vega et al. 2022).

Among the technological applications supporting the observation and monitoring of fishing activities, electronic monitoring has emerged as a cost-efficient tool for reducing discards and bycatch (Van Helmond et al. 2020). However, integrating such tools and electronic reporting, methodological, computational, and connectivity advancements presents new challenges. Chief among these is efficiently handling the vast amounts of data generated while transforming it into actionable information and knowledge for fisheries management (Komoroske & Lewison 2015, Bradley et al. 2019, Gilman et al. 2019, 2020). Conventional data analysis methods often prove inadequate for mining all relevant insights from largescale datasets. Consequently, establishing appropriate protocols and techniques for monitoring and diagnosing catches, discards, and bycatch becomes essential for providing scientific guidance and supporting sustainable management. In this context, emerging deep learning and computer vision techniques offer promising new avenues for automated identification, real-time data processing, and enhanced ecosystem understanding through improved detection of non-target species. These advancements can substantially benefit resource-limited fisheries management strategies by reducing human error, bias, and operational expenses.

Numerous studies have already demonstrated the potential of advanced computer vision methods for automated fish recognition and segmentation. Lu et al. (2020) achieved 96% accuracy using VGGI6 to classify tunids, billfishes, and sharks in deck images. Ju & Xue (2020) reported a 97% accuracy with a customized AlexNet (FAN) for fish species. Similarly, Ovalle et al. (2022) used Mask R-CNN segmentation and MobileNet-V1 measurement to achieve an average accuracy of 98% under low overlapping conditions. Dos Santos & Goncalves (2019) introduced a three-branched Convolutional Neural Networks (CNN) model for 68 fish species, reaching up to 87% accuracy in species-level classification. Villon et al. (2018) and Xu et al. (2021) also documented high accuracies (98% and over 91%, respectively) using CNNs or transfer learning methods. In Japan, Miyazono & Saitoh (2018) employed a transfer learning model based on CIFAR-10 and AlexNet for 50 species, with 71 to 91% accuracy.

Building on these developments, this article proposes an automated deep-learning approach for fish species identification in Chile's central-south pelagic and demersal fisheries. In particular, 18 fish species are classified from self-collected images using CNNs. The objective is to establish a robust methodology demonstrating how these computational methods can advance fisheries management, laying the foundation for a more comprehensive framework. Such a framework would recognize fish species on board and aid in quantifying catches and monitoring discards and bycatch. This integrated system is expected to bolster the effectiveness of Fisheries Observer programs, enhance compliance, and promote sustainable resource use across Chilean fisheries and wider marine ecosystems. Ultimately, the approach aims to improve the estimation of bycatch and discards through real-time detection and quantification, thus supporting more informed mitigation strategies and policy decisions. All code developed in this study is available in our GitHub repository (Alvarado & Plaza-Vega 2023), provided as supplementary material.

MATERIALS AND METHODS

Data

In this research, we utilized a comprehensive dataset composed of 8,118 high-resolution images (each with dimensions of 1,200x800 pixels), featuring 18 species from the central-southern regions of Chile that are objective and non-objetive fisheries (i.e. accompanying fauna), bycatch, and discard species are considered from both pelagic and demersal fisheries. These images were meticulously categorized by expert scientific observers, playing a pivotal role in the continuous monitoring activities within the fisheries sector. For detailed information on the specific species included in this research and the number of images corresponding to each species, please refer to Table 1.

The dataset is characterized by its diversity in background settings and illumination conditions, as illustrated in Figure 1. This study encompasses self-collected images acquired as part of an initiative to advance technological advancement in the monitoring programs of pelagic and demersal fisheries conducted by the IFOP between March 2020 and October 2022.

To optimize the 8,118 images for model development, the dataset was partitioned into training, validation, and testing subsets according to a 70-20-10 ratio. Owing to the limited availability of images for certain species, -Brama australis, in particular (Table 1)-data augmentation techniques (Shorten & Khoshgoftaar 2019) were applied to enhance model generalization and mitigate overfitting. Specifically, the following transformations were incorporated:

* Flipping (horizontal/vertical): random flips simulated the specimens' varying orientations, accounting for different capture angles and handling positions.

* Rotation: images were rotated within a predefined angle range (+15°) to replicate natural inclinations typically observed in field scenarios.

* Shifting: minor horizontal and vertical shifts (up to 10% of the image dimensions) addressed off-center captures commonly occurring during photographic collection.

* Shearing: a moderate 0.2 shearing factor introduced slight skewing, accommodating variations in camera angles.

* Zooming: Small zoom factors (10-20%) reproduced conditions where specimens appear at different scales within the frame.

These transformations effectively expanded and diversified the training subset, enhancing the model's robustness to the inherent variability of fisheries data. To balance computational efficiency with image clarity, each image was resized from 1,200x800 to 500x500 pixels, ensuring sufficient resolution for reliable species identification without unduly increasing computational overhead.

Methodology

Our methodology employs two distinct CNNs (Goodfellow et al. 2016), implemented via the Keras API (Chollet et al. 2015) with a TensorFlow (Abadi et al. 2015) back end, all within the R language environment (R Core Team 2022). Leveraging the computational power of an NVIDIA GeForce RTX 4080 GPU, both models were executed in GPU mode, substantially curtailing the necessary training and testing processing times, thus enhancing the efficiency and feasibility of our approach.

Customized Convolutional Neural Network (CNN)

A customized CNN was designed to identify morphological features in fish species images from central-southern Chilean fisheries. The model (Fig. 2), consists of convolutional layers with 32, 64, 64, 128, and 128 filters, alternated with max-pooling layers, extracting hierarchical features from the images while improving computational efficiency through spatial dimension reduction. Utilizing the rectified linear unit (ReLU) activation function, the model navigates non-linear feature spaces, learning complex patterns from the data. During training, categorical crossentropy and the RMSprop optimizer (Tieleman 2012) (with a learning rate of 10-4) were employed to balance convergence speed and stability. Validation, employing augmented images, ensured model generalization and mitigated overfitting. The final layers use a Softmax activation function, outputting probability distributions across 18 fish species categories. Dropout layers (Srivastava et al. 2014) were incorporated to prevent overfitting, randomly nullifying 25% of neurons during training, enhancing model robustness and predictive capabilities across varied inputs.

Adapted VGG16 Network

The VGG16 model, developed by the Visual Geometry Group at the University of Oxford (Simonyan & Zisserman 2014), was selectively adapted for the specific task of fish species classification in our study. Initially pre-trained on the substantial ImageNet dataset (Russakovsky et al. 2015), we tailored the model to our unique classification context, while specifically focusing our adaptations on its fully-connected top. As depicted in Figure 3, the modifications included the addition of convolutional layers with 64 filters and a 3x3 kernel size, utilizing a ReLU activation function. The implementation of global average pooling followed this, as did the inclusion of dense layers and the application of dropout for regularization. The adaptation culminated in a final dense layer featuring 18 nodes, each representing a fish species, and employed a Softmax activation function to produce probability distributions across the species categories.

RESULTS

Building upon the insights from Alsmadi & Almarashdeh (2022), Barbedo (2022), and Rubbens et al. (2023), and the employment of accuracy as a classification metric is notably prevalent in marine sciences, particularly in image classification applications. This metric's appeal lies in its clear and direct approach to quantifying a model's performance, making it an invaluable marine science tool. In the specific context of fish image classification, where the primary objective is accurately identifying species through visual data, accuracy directly measures a model's effectiveness, reflecting the proportion of correct identifications made.

During the training phase, the custom CNN model demonstrated a training loss near 0.6, with the validation loss slightly lower, which reflects a well-tuned learning process with minimal overfitting, which was further corroborated by the accuracy metrics. Both the training and validation accuracy levels were observed to be above 80%, indicating the model's proficiency in accurately classifying the species, as seen in Figure 4.

In comparison, the VGG16-based pretrained model, adapted to the specificities of the Chilean fish species dataset, showed a refined performance. It registered a training loss of around 0.5 and a notably lower validation loss of approximately 0.4, suggesting a more efficient learning curve and a better generalization capability than the custom CNN model. The accuracy measures were particularly high, with the training and validation accuracy surpassing 80%, as depicted in Figure 4.

The evaluation of the models using confusion matrices 1s shown (Figs. 5-6). The matrices provide a comprehensive view of the classification performance across the different species in the dataset.

The confusion matrices for the custom CNN model (Fig. 5) and the adapted VGG16 model (Fig. 6) are provided to demonstrate the prediction distributions across various classes. Figure 5 shows the classification matrix for the custom CNN model tested on 811 images spanning 18 species, achieving an overall accuracy of approximately 86% (CI: (0.8355; 0.8826]). This model exhibits strong performance for most species but shows lower predictive accuracy for species like Brama australis, Genypterus maculatus, Seriolella punctata, Strangomera bentincki, and Bovichtus chilensis, highlighting arcas for potential improvement. Conversely, the adapted VGG16 CNN model, as depicted in Figure 6, evaluated on the same test set, demonstrates superior performance with an accuracy rate nearing 95% (CI: [0.9355; 0.9651]). Despite its effectiveness, the adapted VGGI6 model faces challenges in accurately classifying certain species, particularly В. australis and S. bentincki, with 33% and 53% classification rates, respectively, suggesting specific targets for future enhancement.

The performance metrics for the two models are provided in Tables 2 and 3. We observed exceptional sensitivity in both models for species such as Caelorinchus fasciatus and Merluccius gayi gayi, with the values reaching 1 and 0.966, respectively, in the custom CNN, and 1 for both species in the VGG16 model. Specificity, indicating the accuracy in identifying true negatives, was uniformly high across most species, often reaching the maximum of 1. Precision and Negative Predictive Value (NPV) metrics, assessing the reliability of the classification of the models, were generally high. However, a notable outlier was the precision for B. australis in the custom CNN model, which was marked as NaN due to a lack of positive predictions. Balanced accuracy, a vital metric in datasets with class imbalances, demonstrated the models' consistent accuracy across various classes, with many species, such as Thyrsites atun and Eleginops maclovinus, achieving scores close to 0.989 and 0.978 in the custom CNN, and 0.991 and 0.999 in the VGG16 model, respectively.

The heatmap visualization is achieved by overlaying the activation map onto the original image, providing an intuitive graphical representation of model focus. Figure 7 presents Gradient-weighted Class Activation Mapping (Grad-CAM) visualization of randomly selected images from each species included in the study.

We employed Grad-CAM to visually interpret the decision-making process of our CNN. Grad-CAM is a widely acknowledged technique that produces visual explanations for decisions from many CNN-based models, rendering it model-agnostic (Selvaraju et al. 2017).

By utilizing the gradients of the target concept, which flow into the final convolutional layer, Grad-CAM generates a coarse localization map that highlights the important regions in the image for predicting the concept. Specifically, it leverages the spatial information preserved in convolutional layers to understand which parts of the input image are deemed significant by the CNN.

DISCUSSION

This work demonstrates the efficacy of deep learning in classifying 18 fish species relevant to pelagic and demersal fisheries in the central-south region of Chile. To achieve this objective, two CNN models, namely custom CNN and adapted VGG16 CNN, were deployed and evaluated regarding their general classification performance and species-specific identification capabilities. The results indicated commendable performance for both models during the training and testing phases, underscoring the effectiveness of CNN-type models in overall fish species classification. The adapted VGG16 CNN model exhibited significantly superior outcomes to the custom CNN model.

Despite the overall effectiveness of the adapted VGG16 model, certain species remain challenging to classify accurately. In particular, B. australis exhibits a low classification rate (33%), largely because it is often misclassified as Thyrsites atun or Stromateus stellatus. Determining the precise morphological features causing this confusion is difficult due to the limited number of B. australis images (30) available for training, which restricts the model's ability to learn distinctive characteristics for this class. S. bentincki is frequently misclassified as E. ringens, reflecting a broader issue of morphological similarity among small pelagic species that compromises the model's discriminative capacity. These findings underscore the importance of augmenting datasets with a more extensive and diverse range of images, thereby providing CNNs with a wider variety of examples to learn. Since CNNs rely on multi-level feature extraction from pixel data, the availability of richer image repositories is crucial for capturing subtle interspecies variations and improving classification performance across all taxa under investigation (Barbedo 2018, Rekha et al. 2020, Yang et al. 2021, Alsmadi & Almarashdeh 2022).

Recent studies have underscored the efficacy of technologies like Remote Electronic Monitoring (REM) in mitigating bycatch in small-scale fisheries, demonstrating how REM can support observer data, enhance accuracy, monitor the effectiveness of mitiga- tion technologies, and guarantee enforcement; this offers a strategic avenue to oversee and curb illegal fishing practices (Worm et al. 2013, Van Helmond et al. 2015, Bartholomew et al. 2018, Glemare et al. 2020, Brown et al. 2021). On that note, advancements in technology, particularly in image recognition, are progressively enhancing the monitoring capabilities of exploitative activities in fisheries. Coupled with decreased data storage costs, these advancements position electronic monitoring as a cost-effective adjunct to fisheries observer programs, promising a substantial shift in operational dynamics.

Barbedo (2022) delineates four pivotal activities in fisheries monitoring: recognition (detecting and quan- tifying individuals), measurement (assessing individual characteristics like weight, size, sex, etc.), tracking (where feasible), and classification (identifying species and pertinent features). Traditional reliance on visual methods, whether direct or via images/videos, encounters challenges including high costs, low throughput, subjectivity, and observer biases, compromising data reliability (Benoit et al. 2009, Cahalan et al. 2016, Snyder & Erbaugh 2020). Additionally, imaging devices facilitate data collection in hazardous environments, mitigating risks to human operators. Furthermore, addressing the knowledge and management gaps related to bycatch necessitates a multi-faceted approach, starting with precisely identifying the fisheries and species affected by bycatch (Soykan et al. 2008, Poisson et al. 2022). This imperative paves the way for the development of targeted technological solutions and research, particularly in species identification, such as the case of this work. For example, the ¿Observer device (Vilas et al. 2020), specifically designed for fishing vessels, automatically captures and processes catch images for species identification and quantification. This system exemplifies the integration of image processing and real-time data analysis, leveraging open-source image recognition software for accurate species detection and quantification, becoming essential in evolving visual tracking systems that autonomously detect, classify, and enumerate various fish species based on video footage from fishing operations. Such integration and recent advancements in computer vision systems highlight the growing feasibility of enhancing electronic monitoring in fisheries, providing comprehensive records of catch and bycatch that outstrip the capabilities of manual observers (Khokker et al. 2022, Lekunberri et al. 2022, da Silva et al. 2023).

Noteworthy is the realization that fishing vessel conditions are seldom ideal, with each fishery having distinct operational procedures and fishing gears of varied selectivities. Challenges like variable lighting, weather conditions, and the need for extensive, annotated datasets necessitate tailored solutions to improve system accuracy and reliability. Thus, accurate identification of individual fish species becomes paramount in environments where multiple species coexist and traditional counting methods often fall short, yielding unreliable or inconsistent data. The CNN-based species classification developed in this study 1s of particular importance as it provides the necessary building blocks for implementing integrated systems that further aid in monitoring fisheries, highlighting the significance of our research in contributing to the broader objectives of sustainable fisheries management and ecosystem conservation. In this context, effective species classification emerges as a critical component for comprehensive surveys and monitoring of fisheries activities. Such classification aids in detecting non-targeted species, facilitates the implementation of effective control measures, and contributes to the reduction of bycatch (Siddiqui et al. 2018, Hridayami et al. 2019, Agarwal et al. 2021). Consequently, integrating these considerations with the advancements in computer vision and electronic monitoring emphasizes the potential for significantly enhancing fisheries management's accuracy, efficiency, and sustainability. Combining rigorous species identification with innovative monitoring technologies, this dual approach addresses the complexities and challenges inherent in modern fisheries management, especially in diverse and dynamic marine ecosystems.

CNN-based models offer state-of-the-art classification accuracy, even when working with limited datasets. However, the manual collection and curation of large image repositories remain significant constraints, underscoring the need for techniques that utilize artificially generated training data. Data augmentation -employing techniques such as flips, rotations, shifts, shearing, and zooming- continues to be one of the most accessible and effective strategies for increasing dataset diversity and improving generalization (Mumuni & Mumuni 2022, Kumar et al. 2024, Wang et al. 2024). These conventional methods can substantially enhance the robustness of models across varied scenarios, particularly in fisheries research, where obtaining comprehensive imagery of underrepresented species can be challenging (BenTamou et al. 2022).

Nevertheless, traditional augmentation approaches have inherent limitations, as they do not fully capture the complexity of real-world conditions or adequately address extreme data imbalances. More advanced techniques, such as generative adversarial networks or class-specific synthetic image generation, could yield higher-quality augmented samples and improve model performance (Mumuni & Mumuni 2022, Kuntalp & Diizyel 2024). However, implementing these sophisticated methods requires specialized, domain-specific procedures. Given that the current study represents an initial effort in this field, extending the scope to include such advanced methods was not feasible. Future research will focus on integrating these more specialized augmentation strategies, particularly for underrepresented and visually complex species. As stated in the Introduction section, the Chilean fisheries administration established the Fisheries Observer and Monitoring Program (FOMP), which incorporates various tasks aimed at achieving the objectives outlined in the new fisheries regulations, redirecting its focus towards the research and monitoring of discard and bycatch in pelagic purse-seine and demersal trawling fisheries. In this context, the FOMP provides information for developing national plans by fishery/fleet to help reduce discards of target and non-target species and reduce bycatch of birds, mammals, and marine reptiles (turtles). The FOMP, with its focus on the conservation and management of hydrobiological resources through data collection, could directly benefit from this research by implementing the proposed models in their sampling and data collection processes, thereby helping to identify different species better and build datasets for the recursive improvement of the implemented models.

In wildlife sciences, emerging methods for data collection are constantly being developed (e.g. unoccupied aircraft systems, camera tags, and satellite imagery). In the fisheries science framework, the latter poses challenges in overcoming the constantly generated flood of information and properly analyzing the data to generate knowledge. To address this issue, there is a growing investment in using artificial intelligence (AI) for automated image data processing. Angliss et al. (2020) outline five key stages of implementing an Al project: scoping, data preparation and annotation, model selection, training, testing, model evaluation and re-training, deployment, and integration. In that sense, future work aims to build upon the good results of this article, integrating them into a broader framework that leverages computer vision to recognize onboard fish, quantify catches, and detect suspicious activities such as discards and bycatch. This comprehensive approach can greatly benefit Fisheries Observer programs, by integrating this automated species classification model into a broader framework that uses computer vision and segmentation techniques for the identification of multiple fish species, automatic measurement of lengths and weights, among other direct applications, thus enhancing compliance and contributing to the sustainable management of Chilean fisheries and broader ecosystems.

Through an examination of the dynamic and transdisciplinary facets of the issue, deep learning methods could be further integrated into more intricate management models, employing automated fish recognition through computer vision systems. Thus, this approach must recognize the interconnectedness of ecological and human dimensions in the context of bycatch, deviating from conventional methodologies and extending beyond immediate benefits for fisheries management. lts potential in accurate bycatch estimation establishes a cornerstone for sustainable practices. This investment in advanced technologies addresses the current complexities of global bycatch. It lays the groundwork for future mitigation efforts, highlighting the transformative power of integrating deep learning and computer vision in marine conservation and fisheries management.

Credit author contribution

E. Alvarado: conceptualization, investigation, methodology, software, formal analysis, validation, writing - original draft; F. Plaza-Vega: conceptualization, investigation, software, formal analysis, visualization, validation, writing - original draft; C. Montenegro: data curation, supervision, writing - review & editing; O. Saavedra: supervision, writing - review & editing. All authors have read and accepted the published version of the manuscript.

Conflict of interest

The authors declare no conflict of interest.

Data availability

The code and data used in this study are available on GitHub (https://github.com/ealvnrz/DL_fish_classifi-cation).

ACKNOWLEDGEMENTS

Eloy Alvarado and Francisco Plaza-Vega were partially supported by Agencia Nacional de Investigacion y Desarrollo, Proyecto ANID/FONDECY T/INICIACION N·11250479 and N°11251120, respectively.

Sidebar

Associate Editor: Sergio Neira

Received: October 16, 2024; Accepted: March 21, 2025

References

REFERENCES

Abadi, M., Agarwal, A., Barham, P., et al. 2015. TensorFlow: Large-scale machine learning on heterogeneous - systems. [https://tensorflow.org]. Reviewed: June 30, 2024.

Agarwal, AK., Tiwari, R.G., Khullar, V., et al. 2021. Transfer learning inspired fish species classification. In 2021 8th International Conference on Signal Processing and Integrated Networks (SPIN), pp. 1154-1159.

Alsmadi, М.К. & Almarashdeh, 1. 2022. A survey on fish classification techniques. Journal of King Saud University-Computer and Information Sciences, 34: 1625-1638.

Alvarado, E. & Plaza-Vega, F. 2023. Deep learning-based classification of species in central-southern fisheries in Chile. [https://github.com/ealvnrz/DL fish classification]. Reviewed: June 30, 2024.

Angliss, R., Sweeney, K., Moreland, E., et al. 2020. Report of the image processing workshop. NOAA Technical Memorandum, NMFS-AFSC: 408 pp.

Barbedo, J.G.A. 2018. Impact of dataset size and variety on the effectiveness of deep learning and transfer learning for plant disease classification. Computers and Electronics in Agriculture, 153: 46-53.

Barbedo, J.G.A. 2022. A review on computer vision and artificial intelligence use for fish recognition, monitoring, and management. Fishes, 7: 335.

Bartholomew, D.C., Mangel, J.C., Alfaro-Shigueto, J., et al. 2018. Remote electronic monitoring as a potential alternative to onboard observers in small-scale fisheries. Biological Conservation, 219: 35-45.

Ben-Tamou, A., Benzinou, A. & Nasreddine, K. 2022. Targeted data augmentation and hierarchical classification with deep learning for fish species identification in underwater images. Journal of Imaging, 8: 214.

Benoit, H.P. & Allard, J. 2009. Can the data from at-sea observer surveys be used to make general inferences about catch composition and discards? Canadian Journal of Fisheries and Aquatic Sciences, 66: 2025-2039.

Bradley, D., Merrifield, M., Miller, К.М., et al. 2019. Opportunities to improve fisheries management through innovative technology and advanced data systems. Fish and Fisheries, 20: 564-583.

Brown, C.J., Desbiens, A., Campbell, M.D., et al. 2021. Electronic monitoring for improved accountability in western Pacific tuna longline fisheries. Marine Policy, 132: 104664.

Cahalan, J., Faunce, C., Bonney, J., et al. 2016. А field test of fisheries observer sampling methods for estimation of at-sea discards. Fisheries Research, 174: 219-233.

Chollet, Е 2015. Keras. [https://github.com/fchollet/ keras]. Reviewed: June 30, 2024.

da Silva, C., Samaai, T., Kerwath, S., et al. 2023. Leaping into the future: Current application and future direction of computer vision and artificial intelligence in marine sciences in South Africa. Research Ideas and Outcomes, 9: : e112231.

Davies, R., Cripps, S.J., Nickson, A., et al. 2009. Defining and estimating global marine fisheries bycatch. Marine Policy, 33: 661-672.

dos Santos, A.A. & Goncalves, W.N. 2019. Improving Pantanal fish species recognition through taxonomic ranks in convolutional neural networks. Ecological Informatics, 53: 100977.

Food and Agriculture Organization (FAO). 1995. Code of conduct for responsible fisheries. FAO, Rome.

Gilman, E., Castejón, V.D.R., Loganimoce, E., et al. 2020. Capability of a pilot fisheries electronic monitoring system to meet scientific and compliance monitoring objectives. Marine Policy, 113: 103792.

Gilman, E., Legorburu, G., Fedoruk, A., et al. 2019. Increasing the functionalities and accuracy of fisheries electronic monitoring systems. Aquatic Conservation: Marine and Freshwater Ecosystems, 29: 901-926.

Glemarec, G., Kindt-Larsen, L., Lundgaard, L.S., et al. 2020. Assessing seabird bycatch in gillnet fisheries using electronic monitoring. Biological Conservation, 243: 108461.

Goodfellow, 1., Bengio, Y. & Courville, A. 2016. Deep learning. MIT Press, Cambridge.

Hridayami, P., Putra, LK.G.D. & Wibawa, K.S. 2019. Fish species recognition using VGG16 deep convolutional neural network. Journal of Computing Science and Engineering, 13: 124-130.

Ju, Z. & Xue, Y. 2020. Fish species recognition using an improved AlexNet model. Optik, 223: 165499.

Kelleher, K. 2005. Discards in the world's marine fisheries: an update. FAO, Rome.

Khokher, M.R,, Little, L.R., Tuck, G.N., et al. 2022. Early lessons in deploying cameras and artificial intelligence technology for fisheries catch monitoring: where machine learning meets commercial fishing. Canadian Journal of Fisheries and Aquatic Sciences, 79: 257-266.

Komoroske, L.M. & Lewison, RL. 2015. Addressing fisheries bycatch in a changing world. Frontiers in Marine Science, 2: 83.

Kumar, T., Brennan, R., Milco, A., et al. 2024. Image data augmentation approaches: A comprehensive survey and future directions. IEEE Access, 12: 187536-187571. doi: 10.1109/ACCESS.2024.3470122

Kuntalp, M. & Diizyel, O. 2024. A new method for GAN-based data augmentation for classes with distinct clusters. Expert Systems with Applications, 235: 121199.

Lekunberri, X., Ruiz, J., Quincoces, I, et al. 2022. Identification and measurement of tropical tuna species in purse seiner catches using computer vision and deep learning. Ecological Informatics, 67: 101495.

Lu, Y.-C, Tung, С. & Kuo, Y.-F. 2020. Identifying the species of harvested tuna and billfish using deep convolutional neural networks. ICES Journal of Marine Science, 77:1318-1329.

Miyazono, T. & Saitoh, T. 2018. Fish species recognition based on CNN using annotated images. IT Convergence and Security, 1: 156-163.

Mumuni, A. & Mumuni, F. 2022. Data augmentation: A comprehensive survey of modern approaches. Array, 16: 100258.

Ovalle, J.C., Vilas, С. & Antelo, L.T. 2022. On the use of deep learning for fish species recognition and quantification on board fishing vessels. Marine Policy, 139: 105015.

Poisson, F., Budan, P., Coudray, S., et al. 2022. New technologies to improve bycatch mitigation in industrial tuna fisheries. Fish and Fisheries, 23: 545-563.

R Core Team. 2022. R: A language and environment for statistical computing. R Foundation for Statistical Computing, Vienna.

Rekha, B., Srinivasan, G., Reddy, S.K., et al. 2020. Fish detection and classification using convolutional neural networks. In: Computational Vision and Bio-Inspired Computing: ICCVBIC 2019. Springer, Berlin, pp. 1221-1231.

Román, C., Escobar, V., San Martín, M., et al. 2022. Informe final programa de investigación del descarte y captura de pesca incidental en pesquerías demersales. Programa de monitoreo y evaluación de los planes de reducción del descarte y de la pesca incidental, año 2021-2022. Reporte técnico. Instituto de Fomento Pesquero, Valparaíso.

Rubbens, P., Brodie, S., Cordier, T., et al. 2023. Machine learning in marine ecology: an overview of techniques and applications. ICES Journal of Marine Science, 80: 1829-1853.

Russakovsky, O., Deng, J., Su, H., et al. 2015. ImageNet large scale visual recognition challenge. International Journal of Computer Vision, 115: 211-252.

Selvaraju, R.R., Cogswell, M., Das, A., et al. 2017. Grad-CAM: Visual explanations from deep networks via gradient-based localization. In: Proceedings of the IEEE International Conference on Computer Vision, pp. 618-626.

Shorten, C. & Khoshgoftaar, TM. 2019. A survey on image data augmentation for deep learning. Journal of Big Data, 6: 1-48.

Siddiqui, S.A., Salman, A., Malik, M.L, et al. 2018. Automatic fish species classification in underwater videos: exploiting pre-trained deep neural network models to compensate for limited labelled data. ICES Journal of Marine Science, 75: 374-389.

Simonyan, K. & Zisserman, A. 2014. Very deep convolutional networks for large-scale image recognition. arXiv:1409.1556. doi: 10.48550/arXiv. 1409.1556

Snyder, H.T. & Erbaugh, J.T. 2020. Fishery observers address Arctic fishery discards. Environmental Research Letters, 15: 0940c4.

Soykan, C.U., Moore, J.E., Zydelis, R., et al. 2008. Why study bycatch? An introduction to the theme section on fisheries bycatch. Endangered Species Research, 5: 91-102.

Srivastava, N., Hinton, G., Krizhevsky, A., et al. 2014. Dropout: a simple way to prevent neural networks from overfitting. Journal of Machine Learning Research, 15: 1929-1958.

Subsecretaría de Pesca y Acuicultura (SUBPESCA). 2022. Informe sectorial de pesca y acuicultura 2018-2022. SUBPESCA, Valparaíso.

Tieleman, T. 2012. Lecture 6.5-rmsprop: Divide the gradient by a running average of its recent magnitude. COURSERA: Neural Networks for Machine Learning, 4: 26-31.

Van Helmond, A.T., Chen, C. & Poos, J.J. 2015. How effective is electronic monitoring in mixed bottom-trawl fisheries? ICES Journal of Marine Science, 72: 1192-1200.

Van Helmond, A.T., Mortensen, L.O., Plet-Hansen, K.S., et al. 2020. Electronic monitoring in fisheries: lessons from global experiences and future opportunities. Fish and Fisheries, 21: 162-189.

Vega, R., Ossa L., Suárez B., et al. 2022. Informe final programa de investigación y monitoreo del descarte y la captura de pesca incidental en pesquerías pelágicas, 2021-2022. Reporte técnico. Instituto de Fomento Pesquero, Valparaíso.

Vilas, C., Antelo, L.T., Martin-Rodríguez, F., et al. 2020. Use of computer vision onboard fishing vessels to quantify catches: The ¡Observer. Marine Policy, 116: 103714.

Villon, S., Mouillot, D., Chaumont, M., et al. 2018. A deep learning method for accurate and fast identification of coral reef fishes in underwater images. Ecological Informatics, 48: 238-244.

Wang, Z., Wang, P., Liu, K., et al. 2024. A comprehensive survey on data augmentation. arXiv:2405.09591. doi: 10.48550/arXiv.2405.09591

Worm, B., Davis, B., Kettemer, L., et al. 2013. Global catches, exploitation rates, and rebuilding options for sharks. Marine Policy, 40: 194-204.

Xu, X., Li, W. & Duan, Q. 2021. Transfer learning and SE-ResNet152 networks-based for small-scale unbalanced fish species identification. Computers and Electronics in Agriculture, 180: 105878.

Yang, X., Zhang, S., Liu, J., et al. 2021. Deep learning for smart fish farming: applications, opportunities and challenges. Reviews in Aquaculture, 13: 66-90.