目的 赤潮是一种常见的海洋生态灾害，严重威胁海洋生态系统安全。及时准确获取赤潮的发生和分布信息可以为赤潮的预警和防治提供有力支撑。然而，受混合像元和水环境要素影响，赤潮分布精细探测仍是挑战。本文针对赤潮边缘探测的难点，结合赤潮边缘高频特征学习与位置语义，提出了一种计算量小、精度高的网络模型（red tide detection network, RTDNet）。方法 针对赤潮边缘探测不准确的问题，设计了基于residual-in-residual（RIR）结构的网络，以提取赤潮边缘水体的高频特征；利用多感受野结构和坐标注意力机制捕获赤潮水体的位置语义信息，增强赤潮边缘水体的细节信息并抑制无用的特征。结果 在GF1-WFV赤潮数据集上的实验结果表明，所提出的RTDNet模型赤潮探测效果不仅优于支持向量机(support vector machine,SVM) 、U-Net、DeepLabv3+、HRNet等通用机器学习和深度学习模型，而且也优于赤潮指数法GF1_RI以及赤潮探测专用深度学习模型RDU-Net，赤潮误提取、漏提取现象明显减少，F1分数在两张测试图像上分别达到了0.905和0.898，相比较于性能第二的模型DeepLabv3+提升了2%以上。而且，所提出的模型参数量小，仅有2.65M，约为DeepLabv3+的13%。结论 本文面向赤潮探测提出一种基于RIR结构的赤潮深度学习探测模型，通过融合多感受野结构和注意力机制提升了赤潮边缘探测的精度和稳定性，同时有效降低了计算量。本文方法展现较好的应用效果，可适用于不同高分辨率卫星影像的赤潮探测。

Objective Red tide is a harmful ecological phenomenon in the marine ecosystem, which seriously threatens the safety of the marine economy. Accurate detection of the occurrence and distribution area of small-scale red tide can provide basic information for the prediction and early warning of red tide. Red tide has the characteristics of short duration and rapid change, and the on-site observation can hardly meet the requirements of timely and accurate detection of red tide. In contrast, remote sensing technology has become an important means of red tide monitoring. However, the traditional method of exponential extraction based on spectral features is easily influenced by the ocean background noise, and it is difficult to select the threshold because the marginal watercolor of the red tide is not obvious. The method based on deep learning can extract the red tide information end-to-end without setting the threshold manually, but the red tide information of low frequency and high frequency are treated equally, which hinders the representation ability of the convolutional neural network. In this paper, aiming at the problem of positioning and identifying small-scale red tide marginal waters, a semantic segmentation method for remote sensing detection of small-scale red tide is proposed by combining high-frequency feature learning of red tide with position semantics. Method The structure of residual-in-residual (RIR) is used to extract the high-frequency characteristics of the red tide marginal waters, and the residual branch is alternately composed of multiple residual groups and multiple receptive fields. Within the residual group, the coordinate attention mechanism and dynamic weight mechanism are introduced to capture the position semantic information of the red tide water body. Multi-receptive field structures are used to capture multi-scale information about red tide water bodies. A small-scale red tide detection network(RTDNet) is constructed, which can enhance the detailed information of red tide marginal waters and suppress useless features. In order to verify the validity of the model, we conduct experiments on the red tide data set of GF1-WFV. Due to the limitation of computer computing resources, these remote sensing images are cropped to the size of 64×64 pixels, and data enhancement operations such as flipping, translating, and rotating are performed on the data. Through the above processing steps, a total of 1050 samples are obtained. We choose Adam as the model optimizer, with the learning rate set to 0.0001, the batch size 2, the quantity of epoch 100 rounds, and the loss function binary cross-entropy. The experiment is carried out under Ubuntu 18.04 operating system, the GPU used is NVIDIA GeForce RTX 2080Ti, and the network model is realized in Python 3.6 with Keras 2.4.0 framework. In addition, the precision (P), recall (R), F1-score (F1), and Intersection over Union (IoU) are comprehensively evaluated to quantitatively analyze the effect of the model. Result The experimental results on the GF1-WFV red tide data set show that the model RTDNet in this paper is superior to SVM, U-Net, DeepLabv3+, HRNet, red tide band exponential method GF1_RI, RD-UNet, and other general or special red tide detection models in both qualitative and quantitative results. From the extraction results, the results from RTDNet are more similar to the ground truth, and the extraction effect of red tide marginal waters is better than other models, with much less phenomenon of false extraction and missing extraction. From the quantitative results, the F1 score and IoU of RTDNet reach 0.905, 0.898, and 0.827, 0.815 respectively on the two test images. Compared with the runner-up DeepLabv3+ in performance, the F1 score of RTDNet is increased by more than 0.02, and the IoU is increased by more than 0.05, but the number of model parameters is only 2.65M, which is 13% of the latter. At the same time, the ablation experiment is carried out, which verifies that each module in the RTDNet network model is helpful to improve the effect of red tide detection. The visualization results of some feature maps in different stages of the network show the gradual refining process of the network to extract red tide. Conclusion In this paper, we propose a red tide small-scale remote sensing detection network model based on the Residual-In-Residual structure, multi-receptive field structure, and attention mechanism, which effectively solves the problem of false extraction and missing extraction caused by the inconspicuous watercolor at the edge of red tide, improves the accuracy and stability of red tide marginal water detection and effectively reduces the calculation load. Experimental results show that RTDNet is superior to other methods and models in detecting small-scale red tide in remote sensing images. This method is suitable for remote sensing accurate location and area extraction of early marine disasters (such as red tide, green tide, and golden tide), and has certain reference significance and applicability for other semantic segmentation tasks with fuzzy edges.