Research on nearshore subaqueous geomorphology stability detection based on few-shot learning
- 1East China Normal University, State Key Laboratory of Estuarine and Coastal Research, Shanghai, China (zd_ren@163.com;hqch@sklec.ecnu.edu.cn; 51163904014@stu.ecnu.edu.cn; 51213904054@stu.ecnu.edu.cn;liuruiqingwh@163.com)
- 2College of Intelligent Information Engineering, Chongqing Aerospace Polytechnic College, Chongqing, China(Nanjing, 210016, China)
- 3Nanjing Center of the China Geological Survey, Nanjing, China(sdkjdxjy@163.com;jiazhengyang@mail.cgs.gov.cn;1528584556@qq.com)
Detecting the stability of nearshore subaqueous geomorphology is a crucial challenge for ensuring early warning and controlling the stability of riverbank slopes. Acquiring nearshore subaqueous geomorphology data using unmanned ship-mounted acoustic multibeam systems is difficult, costly, and time-consuming. Moreover, it is often influenced by weather conditions. The limited availability of nearshore subaqueous geomorphology samples suitable for model training, combined with the high similarity between targets of nearshore unstable geomorphology and the background, poses significant challenges for traditional detection methods. In response to issues such as high similarity in subaqueous geomorphology images, large-scale variations in target size, and a scarcity of samples, this study proposes a nearshore subaqueous geomorphology instability detection framework based on Few-shot learning. Firstly, a feature extraction network is designed, replacing the backbone network with a Swin Transformer network. This network employs a feature pyramid network to extract multi-scale geomorphology features containing global information from the query set, facilitating the fusion of features across deep and shallow layers. Secondly, a weight adjustment module is devised to transform the support set into weight coefficients with class attributes. This adjustment helps in adapting the distribution of geomorphology features for detecting new class objects. Experimental results demonstrate that the proposed detection framework achieves desirable performance in terms of average precision and average recall indicators.
Keywords: Subaqueous Geomorphology; Stability Detection; Few-shot learning
How to cite: Ren, Z., Zhang, P., Cheng, H., Teng, L., Chen, J., Jin, Y., Liu, R., Jia, Z., and Zhang, H.: Research on nearshore subaqueous geomorphology stability detection based on few-shot learning, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-13532, https://doi.org/10.5194/egusphere-egu24-13532, 2024.