Fish Tracking, Counting, and Behaviour Analysis in Digital Aquaculture: A Comprehensive Survey

Reviews in Aquaculture, Journal Year: 2025, Volume and Issue: 17(1)

Published: Jan. 1, 2025

ABSTRACT Digital aquaculture leverages advanced technologies and data‐driven methods, providing substantial benefits over traditional practices. This article presents a comprehensive review of three interconnected digital tasks, namely, fish tracking, counting, behaviour analysis, using novel unified approach. Unlike previous reviews which focused on single modalities or individual we analyse vision‐based (i.e., image‐ video‐based), acoustic‐based, biosensor‐based methods across all tasks. We examine their advantages, limitations, applications, highlighting recent advancements identifying critical cross‐cutting research gaps. The also includes emerging ideas such as applying multitask learning large language models to address various aspects monitoring, an approach not previously explored in literature. identify the major obstacles hindering progress this field, including scarcity datasets lack evaluation standards. To overcome current explore potential multimodal data fusion deep improve accuracy, robustness, efficiency integrated monitoring systems. In addition, provide summary existing available for analysis. holistic perspective offers roadmap future research, emphasizing need standards facilitate meaningful comparisons between promote practical implementations real‐world settings.

Language: Английский

---

ConvFishNet: An efficient backbone for fish classification from composited underwater images

Information Sciences, Journal Year: 2024, Volume and Issue: 679, P. 121078 - 121078

Published: June 19, 2024

Language: Английский

---

Automated fish counting system based on instance segmentation in aquaculture

Expert Systems with Applications, Journal Year: 2024, Volume and Issue: 259, P. 125318 - 125318

Published: Sept. 7, 2024

Language: Английский

---

Semi-supervised and weakly-supervised deep neural networks and dataset for fish detection in turbid underwater videos

Ecological Informatics, Journal Year: 2023, Volume and Issue: 78, P. 102303 - 102303

Published: Sept. 11, 2023

Fish are key members of marine ecosystems, and they have a significant share in the healthy human diet. Besides, fish abundance is an excellent indicator water quality, as adapted to various levels oxygen, turbidity, nutrients, pH. To detect underwater videos, Deep Neural Networks (DNNs) can be great assistance. However, training DNNs highly dependent on large, labeled datasets, while labeling turbid video frames laborious time-consuming task, hindering development accurate efficient models for detection. address this problem, firstly, we collected dataset called FishInTurbidWater, which consists collection footage gathered from waters, quickly weakly (i.e., giving higher priority speed over accuracy) them 4-times fast-forwarding software. Next, designed implemented semi-supervised contrastive learning detection model that self-supervised using unlabeled data, then fine-tuned with small fraction (20%) our FishInTurbidWater data. At next step, trained, novel weakly-supervised ensemble DNN transfer ImageNet. The results show leads more than 20 times faster turnaround time between result generation, reasonably high accuracy (89%). same time, proposed waters (94%) accuracy, still cutting by factor four, compared fully-supervised trained carefully datasets. Our code publicly available at hyperlink FishInTurbidWater.

Language: Английский

---

A data-centric framework for combating domain shift in underwater object detection with image enhancement

Applied Intelligence, Journal Year: 2025, Volume and Issue: 55(4)

Published: Jan. 4, 2025

Abstract Underwater object detection has numerous applications in protecting, exploring, and exploiting aquatic environments. However, underwater environments pose a unique set of challenges for including variable turbidity, colour casts, light conditions. These phenomena represent domain shift need to be accounted during design evaluation models. Although methods have been extensively studied, most proposed approaches do not address inherent In this work we propose data-centric framework combating with image enhancement. We show that there is significant gap accuracy popular detectors when tested their ability generalize new domains. used our compare 14 processing enhancement efficacy improve generalization using three diverse real-world datasets two widely algorithms. Using an independent test set, approach superseded the mean average precision performance existing model-centric by 1.7–8.0 percentage points. summary, demonstrated contribution generalization.

Language: Английский

---

Optimizing remote underwater video sampling to quantify relative abundance, richness, and corallivory rates of reef fish

Coral Reefs, Journal Year: 2025, Volume and Issue: unknown

Published: Jan. 16, 2025

Abstract Remote underwater videos (RUVs) are valuable for studying fish assemblages and behaviors, but analyzing them is time-consuming. To effectively extract data from RUVs while minimizing sampling errors, this study developed optimal subsampling strategies assessing relative abundance, richness, bite rates of corallivorous across eight geographically dispersed reef sites on the Great Barrier Reef in Torres Strait. Analyzing 40 frames per 60-min video yielded precise accurate estimates mean number individuals frame (i.e., MeanCount), with systematic (one every 90 s) proved as effective or better than random sampling, depending survey sites. However, approach underestimated species richness by ~ 40%, missing less common species. For estimating rates, 30 min 15 feeding events were optimal, no significant gains precision accuracy further effort. These enhance standardization process efficiency, reducing time required MeanCount rate nine two times, respectively, compared to full annotation.

Language: Английский

---

Automatic detection, identification and counting of deep-water snappers on underwater baited video using deep learning

Frontiers in Marine Science, Journal Year: 2025, Volume and Issue: 12

Published: Feb. 6, 2025

Deep-sea demersal fisheries in the Pacific have strong commercial, cultural, and recreational value, especially snappers (Lutjanidae) which make bulk of catches. Yet, managing these is challenging due to scarcity data. Stereo-Baited Remote Underwater Video Stations (BRUVS) can provide valuable quantitative information on fish stocks, but manually processing large amounts videos time-consuming sometimes unrealistic. To address this issue, we used a Region-based Convolutional Neural Network (Faster R-CNN), deep learning architecture automatically detect, identify count deep-water BRUVS. Videos were collected New Caledonia (South Pacific) at depths ranging from 47 552 m. Using dataset 12,100 annotations 11 snapper species observed 6,364 images, obtained good model performance for 6 with sufficient (F-measures >0.7, up 0.87). The correlation between automatic manual estimates MaxN abundance was high (0.72 – 0.9), Faster R-CNN showed an underestimation bias higher abundances. A semi-automatic protocol where our supported observers BRUVS footage improved 0.96 counts perfect match (R=1) some key species. This already assist semi-automatically process will certainly improve when more training data be available decrease rate false negatives. study further shows that use artificial intelligence marine science progressive warranted future.

Language: Английский

---

Applying deep learning and the ecological home range concept to document the spatial distribution of Atlantic salmon parr (Salmo salar L.) in experimental tanks

Scientific Reports, Journal Year: 2025, Volume and Issue: 15(1)

Published: Feb. 18, 2025

Abstract Measuring and monitoring fish welfare in aquaculture research relies on the use of outcome- (biotic) input-based (e.g., abiotic) indicators (WIs). Incorporating behavioural auditing into this toolbox can sometimes be challenging because sourcing quantitative data is often labour intensive it a time-consuming process. Digitalization process via computer vision artificial intelligence help automate streamline procedure, gather continuous optimisation assist decision-making. The tool introduced study (1) adapts DeepLabCut framework, based machine learning, to obtain pose estimation Atlantic salmon parr under replicated experimental conditions, (2) quantifies spatial distribution through metrics inspired by ecological concepts home range core area, (3) applies inspect variability around feeding. This proof concept demonstrates potential our methodology for automating analysis behaviour relation including detection, variations within between tanks. impact feeding these patterns also briefly outlined, using 5 days as demonstrative case study. approach provide stakeholders with valuable information how their rearing environment small-scale settings used further development technologies measuring future studies.

Language: Английский

---

Seafloor debris detection using underwater images and deep learning-driven image restoration: A case study from Koh Tao, Thailand

Marine Pollution Bulletin, Journal Year: 2025, Volume and Issue: 214, P. 117710 - 117710

Published: Feb. 20, 2025

Traditional detection and monitoring of seafloor debris present considerable challenges due to the high costs associated with underwater imaging devices complex environmental conditions in marine ecosystems. In response these challenges, this field study conducted Koh Tao, Thailand, proposed an innovative cost-effective approach that leverages super-resolution reconstruction (SRR) technology conjunction optimized object model based on YOLOv8. Super-resolution (SR) images reconstructed by seven SRR models were fed into Seafloor-Debris-YOLO (SFD-YOLO) for detection. RDN achieved highest results a signal-to-noise ratio (PSNR) 41.02 dB structural similarity (SSIM) 95.08 % attained state-of-the-art (SOTA) accuracy mean Average Precision (mAP) 91.2 using RDN-reconstructed magnification factor 4. Additionally, provided in-depth analysis influence factors within process, offering quantitative comparison before after reconstruction, as well comparative evaluation across various algorithms novel pretraining strategy. This survey methods, combined technology, marks advancement monitoring, presenting practical solutions enhance image quality affected enabling precise identification debris.

Language: Английский

---

CNN-Based Optimization for Fish Species Classification: Tackling Environmental Variability, Class Imbalance, and Real-Time Constraints

Information, Journal Year: 2025, Volume and Issue: 16(2), P. 154 - 154

Published: Feb. 19, 2025

Automated fish species classification is essential for marine biodiversity monitoring, fisheries management, and ecological research. However, challenges such as environmental variability, class imbalance, computational demands hinder the development of robust models. This study investigates effectiveness convolutional neural network (CNN)-based models hybrid approaches to address these challenges. Eight CNN architectures, including DenseNet121, MobileNetV2, Xception, were compared alongside traditional classifiers like support vector machines (SVMs) random forest. DenseNet121 achieved highest accuracy (90.2%), leveraging its superior feature extraction generalization capabilities, while MobileNetV2 balanced (83.57%) with efficiency, processing images in 0.07 s, making it ideal real-time deployment. Advanced preprocessing techniques, data augmentation, turbidity simulation, transfer learning, employed enhance dataset robustness imbalance. Hybrid combining CNNs intermediate improved interpretability. Optimization pruning quantization, reduced model size by 73.7%, enabling deployment on resource-constrained devices. Grad-CAM visualizations further enhanced interpretability identifying key image regions influencing predictions. highlights potential CNN-based scalable, interpretable classification, offering actionable insights sustainable management conservation.

Language: Английский

---