Identification of glass eel capture equipment in the Yangtze River estuary based on high-spatial-resolution imagery and an improved YOLOv8 model

文献类型： 外文期刊

第一作者： Zhu, Pengfei

作者： Zhu, Pengfei;Zhou, Weifeng;Zhu, Pengfei

作者机构：

关键词： YOLOv8; Small target detection; Glass eel; Deep learning

期刊名称：ECOLOGICAL INFORMATICS （ 影响因子：7.3； 五年影响因子：7.1 ）

ISSN： 1574-9541

年卷期： 2025 年 89 卷

页码：

收录情况： SCI

摘要： Although eel farming has become an industry, people still cannot achieve large-scale artificial reproduction of eels. Hence, the recruitment of eel for aquaculture can only rely on the capture of natural eel fry, i.e. glass eels. The capture intensity of glass eels is crucial for the sustainable development of natural eel resources, especially for wild eel stocks. The continental shelf of China is an important habitat in the life history of Japanese eel. China's fisheries authorities have adopted a special permit regulation for glass eel capture to control the scale and intensity of these activities. However, the scale of glass eel capture in the Yangtze River Estuary and along the Chinese coast is not fully grasped at the macro level, because of the possibility of poaching by illegal and unreported fishing. To address this problem of monitoring glass eel capture along the coast of China, especially in the Yangtze River Estuary, this study explored a method for identifying and monitoring glass eel capture activities from high spatial resolution satellite image of Jilin-1 by improving YOLOv8 model. The sample dataset was created by data labelling, and split into training, validation, and test sets. To avoid the false detection of small targets, we introduce the asymptotic feature pyramid network to replace the original detection head, and add a detection layer for small targets, which improves the accuracy but increases the parameters and computation volume. Then, C2f module was improved with dual convolutional kernels, and the pooling process was improved by introducing the spatial pyramid pooling fast module with enhanced local attention network. Thereupon the detection speed and accuracy are both improved. So the experiment was carried out based on the sample dataset using the improved YOLOv8 model, which showed that the average precision (mAP@50 %) is 94.8 %, 4.5 % higher than that of the original YOLOv8. The improved model proposed in this article improved localization ability and detection accuracy of tiny targets of capture equipment for glass eel from high-spatialresolution images, and hence the method can be used to monitor glass eel capture activities and evaluate the intensity of glass eel capture.

分类号：

  • 相关文献
作者其他论文 更多>>

微信小程序