基于深度学习的SAR图像海洋涡旋自动检测模型
刘太龙(1998—), 男, 硕士研究生, 主要从事中尺度涡方面的研究。email: 13696350617@163.com |
Copy editor: 殷波
收稿日期: 2024-12-29
修回日期: 2025-02-18
网络出版日期: 2025-02-26
基金资助
国家重点研发计划项目(2022YFC3104900)
国家重点研发计划项目(2022YFC3104905)
国家自然科学基金项目(42176180)
Deep learning-based automatic detection model for ocean eddies in SAR images
Received date: 2024-12-29
Revised date: 2025-02-18
Online published: 2025-02-26
Supported by
National Key Research and Development Program of China(2022YFC3104900)
National Key Research and Development Program of China(2022YFC3104905)
National Natural Science Foundation of China(42176180)
传统基于合成孔径雷达(synthetic aperture radar, SAR)数据的涡旋检测方法需手动设定阈值和特征参数, 操作复杂且难以自动化, 现有深度学习模型在检测过程中也存在较高的漏检和误检情况, 难以满足涡旋检测对精度和效率的要求。为解决这些问题, 文章提出一种基于YOLOv8 (you only look once version 8)的改进模型, 即涡旋检测网络(eddy detection network, EddyDetNet), 以克服上述局限性, 提升检测的准确性和效率。该模型在骨干网络(Backbone)和颈部网络(Neck)中引入自适应特征融合模块(adaptive feature compression module, AFCM)和多尺度特征空间金字塔模块(multi-scale feature spatial pyramid module, MFSP)并优化Neck结构和在头部网络(Head)部分新增小目标检测头, 提升了不同尺度涡旋的检测精度。实验结果表明, EddyDetNet在精确率(precision, p)、召回率(recall, r)和平均精确率均值(mean average precision, mAP)指标上较YOLOv8分别提升了2.4%、3.2%和5.5%, 参数量和运算量减少了38.1%和15.8%。与YOLOv8相比, EddyDetNet在降低计算复杂度和参数量的同时, 保持了较高检测精度, 适用于多目标和复杂背景下的涡旋检测任务。
刘太龙 , 谢涛 , 李建 , 王超 , 张雪红 . 基于深度学习的SAR图像海洋涡旋自动检测模型[J]. 热带海洋学报, 2025 , 44(5) : 65 -76 . DOI: 10.11978/2024242
Traditional eddy detection methods based on SAR data require manual threshold setting and feature parameter tuning, making processes complex and difficult to automate. Existing deep learning models also suffer from high rates of false negatives and false positives, failing to meet the accuracy and efficiency requirements of eddy detection. To address these issues, this paper proposes an improved model based on YOLOv8 (you only look once version 8), named EddyDetNet, to overcome the above limitations and enhance both detection accuracy and efficiency. This model introduces an adaptive feature compression module (AFCM) and a multi-scale feature spatial pyramid module (MFSP) in Backbone and Neck, optimizes the Neck structure, and adds a small target detection head in the Head part, thereby improving the detection accuracy of vortices of different scales. Experimental results show that EddyDetNet outperforms YOLOv8 by 2.4%, 3.2%, and 5.5% in precision (p), recall (r), and mean Average Precision (mAP), respectively, while reducing parameter size and computational complexity by 38.1% and 15.8%. Compared to YOLOv8, EddyDetNet reduces computational complexity and parameter size while maintaining high detection accuracy, making it suitable for eddy detection tasks in multi-target and complex background scenarios.
Key words: oceanic eddies; synthetic aperture radar; deep learning; YOLO; object detection
图1 YOLOv8模型结构图Conv代表卷积层, Concat代表拼接层, Upsample代表上采样层, Detect代表检测头, C2f代表特征融合模块, SPPF代表快速空间金字塔池化层 Fig. 1 YOLOv8 model architecture diagram. Conv represents convolutional layer, Concat represents concatenation layer, Upsample represents upsampling layer, Detect represents detection head, C2f represents Cross Stage Partial Network with 2 convolutions, and SPPF represents Spatial Pyramid Pooling Fast layer |
图2 EddyDetNet模型结构图Conv代表卷积层, AFCM代表自适应特征融合模块, MFSP代表多尺度特征空间金字塔模块, Concat代表拼接层, Upsample代表上采样层, Detect代表检测头 Fig. 2 EddyDetNet model architecture diagram. Conv represents convolutional layer, AFCM represents adaptive feature compression module, MFSP represents multi-scale feature spatial pyramid module, Concat represents concatenation layer, Upsample represents upsampling layer, and Detect represents detection head |
图3 自适应特征融合模块(AFCM)结构图Conv代表卷积层, Split代表分割层, PMBlock代表使用PConv和MLCA构建的新模块, Concat代表拼接层 Fig. 3 Structure diagram of AFCM. Conv represents convolutional layer, Split represents split layer, PMBlock represents new module constructed using PConv and MLCA, and Concat represents concatenation layer |
图5 多尺度特征空间金字塔模块(MFSP)结构图Conv代表卷积层, MaxPool2d代表二维最大池化层, EMA代表高效多尺度注意力机制, Concat代表拼接层 Fig. 5 MFSP structure diagram. Conv represents convolutional layer, MaxPool2d represents two-dimensional maximum pooling layer, EMA represents Efficient Multi-scale Attention mechanism, Concat represents concatenation layer |
表1 消融实验Tab. 1 Ablation experiment |
改进方法 | AFCM | MFSP | Neck+Head | p/% | r/% | mAP@0.5/% | 参数量/M | GFLOPs |
---|---|---|---|---|---|---|---|---|
1 | 88.2 | 88.6 | 89.5 | 24.64 | 78.7 | |||
2 | √ | 86.2 | 92.5 | 94.7 | 15.85 | 50.2 | ||
3 | √ | 88.0 | 89.2 | 92.5 | 24.84 | 80.2 | ||
4 | √ | 88.9 | 91.0 | 93.4 | 24.16 | 99.6 | ||
5 | √ | √ | √ | 90.6 | 91.8 | 95.0 | 15.26 | 66.3 |
注: “√”表示实验结果加入了对应模块, 空白表示实验结果未加入对应模块 |
表2 算法性能对比Tab. 2 Algorithm performance comparison |
模型名称 | AP/%(AE) | AP/%(CE) | p/% | r/% | mAP@0.5/% |
---|---|---|---|---|---|
SSD | 47.01 | 81.24 | 58.48 | 56.90 | 64.12 |
Faster R-CNN | 32.99 | 79.01 | 43.67 | 73.76 | 56.00 |
YOLOv7 | 42.60 | 66.30 | 51.90 | 61.10 | 54.50 |
YOLOv8 | 85.20 | 93.70 | 88.20 | 88.60 | 89.50 |
YOLOv9 | 39.70 | 78.80 | 52.30 | 76.00 | 59.20 |
EddyDetNet | 94.20 | 95.80 | 90.60 | 91.80 | 95.00 |
图6 不同模型涡旋在4幅SAR图像上的涡旋检测结果a. SSD; b. Faster R-CNN; c. YOLOv7; d. YOLOv8; e. YOLOv9; f. EddyDetNet。图中蓝色方框处是被检测出来的涡旋, 数字表示模型检测的置信度。CE为气旋涡, AE为反气旋涡 Fig. 6 Eddy detection results of different models on four SAR images. (a) SSD; (b) Faster R-CNN; (c) YOLOv7; (d) YOLOv8; (e) YOLOv9; (f) EddyDetNet |
[1] |
陈戈, 杨杰, 田丰林, 等, 2021. 海洋涡旋遥感: 进展与挑战[J]. 遥感学报, 25(1): 302-322.
|
[2] |
董昌明, 2015. 海洋涡旋探测与分析[M]. 北京: 科学出版社.
|
[3] |
甘滢晖, 2022. 基于遥感资料的中尺度涡与叶绿素分布的关系研究[D]. 大连: 大连海洋大学.
|
[4] |
高山, 王凡, 李明悝, 等, 2007. 中尺度涡的高度计资料同化模拟[J]. 中国科学(D辑: 地球科学), 37(12): 1669-1678.
|
[5] |
沈飙, 陈扬, 杨琛, 等, 2020. 海洋科学中尺度涡的计算机视觉检测和分析方法[J]. 数据与计算发展前沿, 2(6): 30-41.
|
[6] |
吴进群, 陈戈, 马纯永, 等, 2023. 深度学习的SAR图像海洋涡旋自动检测及其特征提取[J]. 舰船科学技术, 45(6): 170-173.
|
[7] |
赵文涛, 俞建成, 张艾群, 等, 2016. 基于卫星测高数据的海洋中尺度涡流动态特征检测[J]. 海洋学研究, 34(3): 62-68.
|
[8] |
郑晓莉, 董庆, 樊星, 2020. 北太平洋中尺度涡海表温度和叶绿素浓度特征分析[J]. 遥感学报, 24(1): 85-96.
|
[9] |
|
[10] |
|
[11] |
|
[12] |
|
[13] |
|
[14] |
|
[15] |
|
[16] |
|
[17] |
|
[18] |
|
[19] |
|
[20] |
|
[21] |
|
[22] |
|
[23] |
|
[24] |
|
[25] |
|
[26] |
|
[27] |
|
[28] |
|
[29] |
|
/
〈 |
|
〉 |