ABSTRACT Offshore aquaculture plays an important role in China’s marine fishery economy. The research on the extraction of offshore aquaculture areas based on remote sensing technology is of great significance for the regulation of offshore fishery resources and the protection of the marine ecological environment. This paper uses the Gaofen-2 series multispectral remote sensing image to extract the offshore aquaculture areas of Lianyungang City. We use the optimum index factor to extract the spectral features of the aquaculture areas and the grey-level co-occurrence matrix to extract their texture features. We use the Bhattacharyya distance to select the spatial and spectrum features and construct the characteristic data set of the aquaculture areas. In this paper, we propose a method to construct a uniform distributed disturbance term to optimize the cross entropy loss function. We employ it in the three-dimensional convolutional neural network (3D-CNN) model, extract the extended feature data set of aquaculture areas, and input it into the radial basis function support vector machine (RBF-SVM) classifier for classification. Within the study area of 150 km2, the experimental results show that the extraction model has high extraction accuracy and strong spatial migration despite complex water backgrounds. The F 1-score values in the training area and the four random test areas were 0.939 or above for 2017 data. In addition, the extraction model also has stable time migration. We used the extraction model on remote sensing data for the same study area in 2018 and 2019. The F 1-scores for all test areas are 0.866 or higher. Therefore, the model proposed in this paper is suitable for the extraction of large-scale and multi-temporal offshore raft aquaculture areas from remote sensing images.

中文翻译：

---

基于3D-CNN模型精确提取近海筏式养殖区

摘要 近海养殖在我国海洋渔业经济中占有重要地位。基于遥感技术的近海养殖区提取研究对近海渔业资源调控和海洋生态环境保护具有重要意义。本文利用高分2号系列多光谱遥感影像提取连云港市近海养殖区。我们使用最优指标因子提取水产养殖区域的光谱特征，使用灰度共生矩阵提取其纹理特征。我们使用 Bhattacharyya 距离选择空间和光谱特征并构建水产养殖区的特征数据集。在本文中，我们提出了一种构建均匀分布扰动项的方法来优化交叉熵损失函数。我们将其应用于三维卷积神经网络（3D-CNN）模型，提取水产养殖区的扩展特征数据集，输入径向基函数支持向量机（RBF-SVM）分类器进行分类。在150 km2的研究区域内，实验结果表明，尽管水体背景复杂，但提取模型提取精度高，空间迁移能力强。2017年数据在训练区和四个随机测试区的F 1-score 值为0.939或以上。此外，提取模型还具有稳定的时间偏移。我们对2018年和2019年同一研究区的遥感数据使用了提取模型。所有测试区的F 1-scores为0。866 或更高。因此，本文提出的模型适用于从遥感影像中提取大尺度、多时相近海筏式养殖区。