Page 111 - 《渔业研究》2025年第6期

802 渔 业 研 究 第 47 卷

Study on remote sensing extraction of coastal aquaculture
ponds based on Res-PGAUnet

1*
1
CHEN Hongmei ，PENG Jun 1,2,3 ，CHEN Yunzhi 2,3* ，LUO Donglian ，
1
2,3
CHEN Yumei ，LIU Guoxin ，WANG Wanping 1
(1. Fisheries Research Institute of Fujian, Xiamen 361013, China;
2. Key Lab of Spatial Data Mining & Information Sharing of Ministry of Education, Fuzhou University, Fuzhou 350108, China;
3. National and Local Joint Engineering Research Center for Comprehensive Application of Satellite and Spatial Information
Technology, Fuzhou 350108, China)

Abstract: [Background] The coastal aquaculture ponds are often mixed with salt fields and river channels, and
the ponds have different forms and scales, which make the traditional remote sensing extraction methods face
the technical bottlenecks such as insufficient extraction accuracy, weak anti-interference ability and low degree
of automation. Deep learning methods, however, can automatically learn rich spectral and spatial features from
images through convolutional layers, enabling large-scale precise classification and enhancing the automation of
extraction tasks. [Objective] This study aimed to realize accurate and efficient automated extraction of
aquaculture ponds in complex interference scenarios. [Methods] This study utilizes domestic GF-2 high-resol-
ution remote sensing imagery data. Building upon the U-Net model, we constructed the Res-PGAUnet model
for the coastal pond aquaculture zone in the south of Jiuzhen Bay in Zhangzhou City, Fujian Province. The mod-
el integrates residual structure, pyramid pooling, guided branches, and dual attention mechanisms, with preci-
sion analysis and large-scale application testing conducted. [Results] Core improvement modules (Residual
structure, pyramid pooling, guided branches, and dual attention mechanism) significantly enhanced perform-
ance. Their combined effect enables Res-PGAUnet to demonstrate stronger anti-interference capability and ro-
bustness when handling diverse interference objects such as rivers, salt pans, and seawater. The IoU and F1-
score reached 0.854 0 and 0.921 3 respectively, effectively reducing false positives and negatives while address-
ing small target omissions and boundary adhesion issues. [Conclusion] Large-scale generalization tests further
validate the practical potential of Res-PGAUnet. The model provides reliable technical support for precise mon-
itoring of pond aquaculture spatial information and sustainable fishery development.
Key words: aquaculture pond; GF-2 imagery; deep learning; Res-PGAUnet; pyramid pooling; guide branches;
dual attention mechanism