Abstract: Raft aquaculture is a key method for farming aquatic products, and utilizing remote sensing technology for its accurate monitoring is an essential approach for controlling aquaculture density and supporting scientific planning. However, because the spatial distribution of raft aquaculture is heavily influenced by waves and specular reflection, significant information loss occurs in many raft aquaculture areas on remote sensing images. To improve extraction precision and accuracy, this paper proposes an extraction method based on the improved Criminisi and SAM (segment anything model) algorithms specifically for raft aquaculture elements with information loss. First, the dark channel prior is used to locate and classify the areas affected by information loss. The Criminisi algorithm is then improved by optimizing priority calculation, sample block generation, and matching criteria to effectively restore the information loss zones. Finally, the SAM algorithm is applied to achieve zero-shot extraction of raft aquaculture elements. In addition, experiments using a Gaofen-6 image dataset were conducted to compare the proposed method with two advanced deep learning-based algorithms, RSMamba (remote sensing mamba) and Swin Transformer. The experimental results demonstrate that the proposed algorithm achieves both the accuracy rate and completeness rate of at least 90%, with an error rate not exceed 0.02%, indicating superior performance.

Key words: Criminisi, SAM, information loss, raft aquaculture, remote sensing extraction