STAFDD: A spatio-temporal automatic fish disease detection method
Bo Wang, Zhiyong Cao, Kailong Yan, Danying Cao, Qi Fang, Wanting Zhang, Yingyin Cheng, Yongming Li 等 11 位
Chinese Academy of Sciences Institute of Hydrobiology Ministry of Agriculture and Rural Affairs University of Chinese Academy of Sciences
阅读操作
确认中在文库中上传 PDF 后可生成中文音频讲解。
摘要与影响
Fish diseases spread rapidly in high-density aquaculture environments, making timely and accurate diagnosis difficult using traditional manual inspection. Existing automated detection methods mainly rely on computer vision to identify diseased fish based on body-surface changes, whereas a smaller number of methods use differences in swimming trajectories for disease detection; both types of approaches have their own advantages and limitations. In this study, we propose a spatio-temporal automatic fish disease detection method, termed STAFDD, which integrates spatial body-surface features with temporal trajectory-based behavioral information to improve the accuracy and stability of diseased fish recognition. The proposed framework first employs an improved YOLOv8 detector, SSCA-YOLO, to perform fish body detection and body-surface abnormality recognition. By incorporating the SPD-CBS downsampling module, a small-object detection head, and a Coordinate Attention (CA) mechanism, SSCA-YOLO improves the localization accuracy of fish targets and body-surface abnormalities. Subsequently, the ByteSort tracking algorithm, which integrates ReID appearance features, is used to achieve stable multi-object tracking, generate individual trajectories, and reduce identity switches. An LSTM temporal model is then applied to extract abnormal behavioral patterns from trajectory and motion features. Finally, a lightweight learnable MLP gating strategy is introduced to adaptively fuse the outputs of the visual and trajectory branches based on the visual confidence, trajectory confidence, and normalized trajectory length, thereby determining the health status of each fish. A fish disease detection dataset was constructed using Aeromonas hydrophila -challenged Prussian carp ( Carassius gibelio ) as the experimental model for training and validation. Experimental results showed that, compared with the original YOLOv8, SSCA-YOLO improved mAP@[0.5:0.95] by 1.58%. The ByteSort tracking algorithm achieved 89.74% HOTA and 94.05% MOTA, with only 37 identity switches. The STAFDD model based on learnable MLP-gated fusion achieved an overall classification accuracy of 96.86%, outperforming both the visual-only branch and the trajectory-only branch. By integrating fish appearance features with temporal behavioral cues, the proposed method provides an effective automated solution for fish disease monitoring in high-density aquaculture environments.
逐年被引趋势
暂无年度引用数据
关键指标
同类平均 = 1
同领域 · 同年份 · 同类型
Google Scholar 与 OpenAlex 的被引统计范围不同,数值存在差异属正常。
AI 辅助阅读
依据:摘要
可就本文提问;依据不足时会说明。
学术脉络
学科主题
物理Water Quality Monitoring Technologies
Smart Agriculture and AI · Aquaculture disease management and microbiota