基于深度学习RCF模型的三都澳筏式养殖区提取研究

Extracting raft aquaculture areas in Sanduao from high-resolution remote sensing images using RCF

三都澳是中国重要的海水养殖海湾,在水产养殖中占有较高的经济地位。快速准确地获取养殖区的分布范围、数量以及面积等信息,对养殖区规划、产值估计、生态调查、风暴潮灾害预防等具有重要的意义。然而,随着养殖区域的扩大,海水背景状态愈发复杂多样,光谱特征差异较大,为养殖区提取带来困难。在本实验中,利用高分辨率遥感卫星GF-2图像,引入深度学习RCF(Richer Convolutional Features)网络模型对海湾内的筏式养殖区进行了提取。结果显示:该方法无需事先对区域进行水陆分离处理,且对水中泥沙较多的区域以及海浪较大的区域有很好的提取效果,提取精度达93%以上,适合进行大规模海水养殖区提取应用。

Sanduao is an important sea-breeding bay in China and holds a high economic status in aquaculture. Quickly and accurately obtaining information such as the distribution area, quantity, and aquaculture area is important for breeding area planning, production value estimation, ecological survey, and storm surge prevention. However, as the aquaculture area expands, the seawater background becomes increasingly complex and spectral characteristics differ dramatically, making it difficult to determine the aquaculture area. In this study, we used a high-resolution remote-sensing satellite GF-2 image to introduce a deep-learning Richer Convolutional Features network model to extract the raft aquaculture area in the bay. The results demonstrate that this method does not require land and water separation of the area in advance, and good extraction can be achieved in areas with more sediment and waves, with an extraction accuracy >93%, which is suitable for large-scale raft aquaculture area extraction.