Meta-Transfer Learning with Attention for Marine Aquaculture Information Extraction

Floating raft aquaculture is a crucial marine aquaculture method, the difference of floating raft aquaculture in different regions in remote sensing images brings challenges to marine aquaculture monitoring by deep learning information extraction. Meta-transfer learning with attention (MTLA) approach is proposed in this paper. Before the meta-training phase, a data enhancement module is included in the network framework to address the problem of few-task scenarios in meta-learning. Transfer attention block designed specifically for extracting information from SAR images by combining the strengths of transfer learning and attention mechanisms to enhance the capability of feature extraction. Actual GF-3 SAR images from different areas are utilized to verify the effectiveness of the proposed methods.