Analysis of Deep Learning Models for Underwater Species Detection and Classification

Abstract

Underwater fish detection is challenging. Fish can be very small in images. The contrast in scenes is often weak. Turbid water can hide details. When fish overlap it becomes harder to locate each one. This study compares deeplearning object detectors using an underwater fish dataset. The focus is on both detection quality and how well the models work in video processing. The models tested are YOLO26, RF-DETR and SSD300 with a backbone. In the evaluations YOLO26s achieved a mAP@50 of 0.875 and a mAP@50:95 of 0.499. RF-DETR Base reached 0.4465 and 0.3085 respectively. The SSD300 model showed a precision of 0.8611, recall of 0.1975, mAP@50 of 0.1436 and mAP@50:95 of 0.0607. These results show that high precision alone does not mean detection performance if recall is low. The comparison also looks at class-label consistency and differences, in settings. The model runs were not all done under the conditions. Still the results give a view of what each detector can do. They highlight the strengths and weaknesses of YOLO26, RF-DETR and SSD300 for detecting fish in underwater.

Introduction

The project focuses on developing and evaluating an automated fish detection system for underwater images and videos using deep learning-based object detection models. The main objective is to accurately identify fish species, locate fish, and process underwater videos at a practical speed for ecological surveys, aquaculture monitoring, and automated fish classification.

Underwater fish detection faces several challenges, including poor visibility, low contrast, colour distortion, overlapping fish, small objects, and similarities between fish species. The study compares three object detection models: YOLO26, RF-DETR, and SSD300, to examine their detection accuracy, localization performance, and processing speed.

The system processes underwater videos by dividing them into individual frames. Each frame is analyzed by a trained model that predicts fish locations using bounding boxes, identifies their classes, and assigns confidence scores. Model performance is evaluated using precision, recall, F1-score, mean Average Precision (mAP@50 and mAP@50:95), and frames per second (FPS).

The experimental results show that YOLO26 achieved the strongest overall detection performance among the evaluated runs, with 84.5% precision, 83.9% recall, 87.5% mAP@50, and 49.9% mAP@50:95 at approximately 17.7 FPS. RF-DETR achieved 44.65% mAP@50 and 30.85% mAP@50:95, while SSD300 recorded 86.11% precision but only 19.75% recall and 14.36% mAP@50. These results indicate that SSD300 produced relatively accurate positive predictions but missed many actual fish.

The study concludes that YOLO26 is a promising option for balancing fish detection accuracy and processing speed in the evaluated experiments. However, the models were tested under different hardware and configuration settings, so the results do not establish a definitive architectural ranking. The project also clarifies that counting unique fish in videos requires an additional object-tracking mechanism to avoid counting the same fish across multiple frames.

In conclusion, the project demonstrates the potential of deep learning for automated underwater fish detection and species classification. It highlights the importance of evaluating accuracy, recall, localization quality, and processing speed together to develop reliable systems for real-world underwater monitoring.

Conclusion

This study looked at YOLO26, RF-DETR and SSD300 using the underwater fish detection data. Among the models tested YOLO26 showed the mAP@50 and also had the fastest processing speed. RF-DETR performed better than SSD in terms of mAP@50:95, which means it did a job when the localization requirements were more strict. SSD300 had a precision value of 0.8611 but its recall was very low at 0.1975. This led to a mAP@50 of 0.1436 and an even lower mAP@50:95 of 0.0607. The confusion matrix and class-wise results show that SSD300 missed detections and had uneven performance across different classes. The findings suggest that YOLO26 is currently the practical choice for real-time underwater video analysis. RF-DETR gives a comparison focused on accuracy. SSD300 serves as a baseline and helps highlight the difference between precision and recall in this specific task. These conclusions are limited to the data and conditions used in this study. A final decision about which model to deploy should only be made after rerunning all models under the evaluation protocol and testing the full detection, to-tracking counting pipeline. Future work should aim to improve detection of fish and fish that are partially hidden in low-visibility underwater settings. A bigger varied dataset could also help classify fish that look very similar. Future steps might add real-time tracking and counting of fish from videos and also test lightweight models to speed up processing on devices, with limited resources.

References

[1] C. Shah et al., “YOLOv8-TF: Transformer-Enhanced YOLOv8 for Underwater Fish Species Recognition with Class Imbalance Handling,” Sensors, vol. 25, no. 6, 1846, 2025. https://doi.org/10.3390/s25061846 [2] I. Robinson, P. Robicheaux, M. Popov, D. Ramanan, and N. Peri, “RF-DETR: Neural Architecture Search for Real-Time Detection Transformers,” in Proc. ICLR, 2026. https://arxiv.org/abs/2511.09554 [3] W. Liu et al., “SSD: Single Shot MultiBox Detector,” in Proc. ECCV, 2016. https://arxiv.org/abs/1512.02325 [4] Research paper comparison slides, “Underwater Fish Detection Model Comparison,” version including RF-DETR and later YOLO-family model discussion, local research paper file, 2026. [5] Local training records for YOLO26s, RF-DETR Base, and the custom YOLOv10 small-dataset run, including results.csv, args.yaml, metrics.csv, and training_config.json, 2026. [6] YOLO26s 1280-pixel run output directory, including results.png, confusion matrices, and validation mosaics, local research paper files, 2026. [7] G. Jocher, J. Qiu, M. Liu, S. Lyu, F. C. Akyon, and M. E. Kalfaoglu, “Ultralytics YOLO26: Unified Real-Time End-to-End Vision Models,” arXiv:2606.03748, 2026. https://arxiv.org/abs/2606.03748 [8] Ultralytics, “YOLO26: Real-Time Vision Models,” Ultralytics Documentation, 2026. https://docs.ultralytics.com/models/yolo26

Copyright

Copyright © 2026 Kamali B, Kishore A, Harshavardhan T, Sakthivel Murugan S. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.