Contrastive Learning Regularization Method for Information Extraction of Floating Raft Aquaculture

Addressing the challenge of coherent speckle noise in synthetic aperture radar (SAR) imagery, as well as the impact of intra-class variability and inter-class similarities on the performance of raft aquaculture information extraction, this paper proposes a contrastive learning regularized triplet attention (CLRTA) approach. Following feature extraction by the encoder, the triplet attention mechanism is incorporated to filter out the noise and focus on the floating raft regions, thereby enhancing the model’s capability to process information in complex backgrounds. Subsequently, a contrastive learning loss function is introduced to minimize distances within the same category while increasing distances between different categories, thus enhancing the model’s discriminative ability towards intra-class variability and inter-class similarities. Information extraction is conducted in the raft aquaculture areas of Shangchuan Island, Guangdong Province, validating the effectiveness of the proposed method.