Machine learning-based fusion technique for sea surface temperature in the Bohai and Yellow Seas
-
摘要: 基于搭载于相应卫星平台的MODIS和AMSR2传感器观测的夜间海表温度（Sea Surface Temperature，SST）数据，通过构建反向传播神经网络（BPNN）、随机森林（RF）和卷积神经网络（CNN）3种机器学习模型，开展海表温度数据融合技术研究。在模型输入设计层面，提出两种差异化方案：基础方案仅包含经纬度与原始 SST 数据，增强方案则在此基础上引入时间参数，进而形成 BP_without_time、BP_with_time、RF_without_time、RF_with_time、CNN_without_time、CNN_with_time 共 6 种融合方案。实验测试结果显示，在3种机器学习模型中，CNN展现出最为优异的性能表现，而RF模型的性能相对较弱。在3种模型的对比测试中，融入时间参数的增强方案均显著优于未包含时间参数的基础方案。基于2023−2024年浮标实测数据的验证结果表明，融合后的 SST 数据精度略低于AMSR2_SST，但相较于MODIS_SST，其精度提升效果显著。融合数据的逐月覆盖率较原始数据大幅提升，2023−2024年最低覆盖率从MODIS的19.79%、AMSR2的32.10% 提升至49.56%以上。同时，高分辨率融合结果能捕捉更精细的温度分布特征，较10 km分辨率的AMSR2数据提供更丰富的空间细节。Abstract: This study is based on sea surface temperature (SST) data from MODIS and AMSR2 satellite observations. Three machine learning models—backpropagation neural network (BPNN), random forest (RF), and convolutional neural network (CNN)—were constructed to conduct research on SST data fusion technology. In terms of model input design, two differentiated schemes are proposed: the basic scheme includes only latitude, longitude, and raw SST data, while the enhanced scheme introduces a time parameter, resulting in six fusion schemes—BP_without_time, BP_with_time, RF_without_time, RF_with_time, CNN_without_time, and CNN_with_time. Experimental test results show that among the three machine learning models, CNN demonstrates the most outstanding performance, while the RF model performs relatively weakly. In comparative tests of the three models, the enhanced schemes incorporating time parameters significantly outperform the basic schemes without time parameters. Validation results based on 2023–2024 buoy measurement data indicate that the accuracy of the fused SST data is slightly lower than that of AMSR2_SST, but shows a significant improvement compared to MODIS_SST. The monthly coverage of the fused data has been significantly improved compared to the original data. The minimum coverage for 2023–2024 increased from 19.79% of MODIS and 32.10% of AMSR2 to over 49.56%. Additionally, the high-resolution fused results can capture more detailed temperature distribution characteristics, providing richer spatial details compared to the 10 km resolution AMSR2 data.
-
Key words:
- sea surface temperature fusion /
- machine learning /
- BPNN /
- RF /
- CNN
-
图 4 6种融合方案在测试集上的验证结果散点图（N表示测试点数）
Fig. 4 Scatter plot of validation results for six fusion schemes on the test set (N represents the number of test points)
图 5 浮标SST对原始MODIS_SST（a）、AMSR2_SST（b）的检验评估结果散点图（2023–2024年，N表示匹配点数）
Fig. 5 Scatter plot of the validation results of buoy SST for original MODIS_SST (a) and AMSR2_SST (b) (2023–2024, N represents the number of matching points)
图 6 6种融合方案SST的浮标检验结果散点图（2023–2024年，N表示匹配点数）
Fig. 6 Scatter plot of buoy verification results for six fusion schemes SST (2023–2024, N represents the number of matching points)
图 7 MODIS_SST、AMSR2_SST和融合SST三者的2023年和2024年逐月覆盖率
Fig. 7 Monthly coverage of MODIS_SST, AMSR2_SST, and fused SST for 2023 and 2024
图 8 2024年2月10日的MODIS_SST、AMSR2_SST和融合SST对比
红框处表示融合SST显示出更丰富的空间分布变化特征
Fig. 8 Comparison of MODIS_SST, AMSR2_SST, and fused SST on February 10, 2024
The red box indicates that the fused SST exhibits richer characteristics of spatial distribution variation
图 9 2024年5月12日的MODIS_SST、AMSR2_SST和融合SST对比
红框处表示融合SST显示出更丰富的空间分布变化特征
Fig. 9 Comparison of MODIS_SST, AMSR2_SST, and fused SST on May 12, 2024
The red box indicates that the fused SST exhibits richer characteristics of spatial distribution variation
图 10 2024年8月28日的MODIS_SST、AMSR2_SST和融合SST对比
红框处表示融合SST显示出更丰富的空间分布变化特征
Fig. 10 Comparison of MODIS_SST, AMSR2_SST, and fused SST on August 28, 2024
The red box indicates that the fused SST exhibits richer characteristics of spatial distribution variation
图 11 2024年10月24日的MODIS_SST、AMSR2_SST和融合SST对比
红框处表示融合SST显示出更丰富的空间分布变化特征
Fig. 11 Comparison of MODIS_SST, AMSR2_SST, and fused SST on October 24, 2024
The red box indicates that the fused SST exhibits richer characteristics of spatial distribution variation
图 12 研究区域内直接合并的SST与浮标对比
Fig. 12 Comparison of directly merged SST and buoy data within the study area
表 1 红外辐射计和微波辐射计观测SST的区别
Tab. 1 Differences between SST observations by infrared and microwave radiometers
传感器类型 工作波段 测量深度 优势 局限性 红外辐射计 热红外
（8～14 μm）表皮层
（几微米）空间分辨率高 受云层、水汽干扰 微波辐射计 微波
（6～37 GHz）次表层
（1～5 mm）穿透云层 分辨率较低 
下载: 导出CSV
表 2 6种融合方案在测试集上的验证结果
Tab. 2 Validation results of six fusion schemes on the test set
融合方案 Std/℃ MAE/℃ MSE/℃ RMSE/℃ Bias/℃ R BP（不包含时间维度） 0.38 0.32 0.15 0.38 −0.06 0.9987 BP（包含时间维度） 0.38 0.32 0.14 0.38 0.04 0.9988 RF（不包含时间维度） 0.47 0.38 0.22 0.47 −0.02 0.9983 RF（包含时间维度） 0.43 0.35 0.19 0.43 −0.04 0.9985 CNN（不包含时间维度） 0.37 0.31 0.14 0.38 −0.06 0.9988 CNN（包含时间维度） 0.34 0.28 0.12 0.35 −0.03 0.9990 注：Std、MAE、MSE、RMSE、Bias数据仅保留2位小数。
下载: 导出CSV
表 3 原始MODIS_SST和AMSR2_SST以及6种融合方案SST的浮标检验结果（2023–2024年）
Tab. 3 Buoy validation results of original MODIS_SST, AMSR2_SST and SST from six fusion schemes (2023–2024)
Std/℃ MAE/℃ MSE/℃ RMSE/℃ Bias/℃ R MODIS_SST 0.69 0.48 0.51 0.71 −0.19 0.9966 AMSR2_SST 0.54 0.46 0.36 0.60 0.24 0.9979 BP_without_time_SST 0.65 0.58 0.43 0.66 −0.06 0.9970 BP_with_time_SST 0.63 0.58 0.40 0.63 0.04 0.9970 RF_without_time_SST 0.70 0.56 0.49 0.70 0.03 0.9967 RF_with_time_SST 0.69 0.57 0.48 0.69 0.03 0.9966 CNN_without_time_SST 0.66 0.58 0.39 0.63 − 0.0014 0.9971 CNN_with_time_SST 0.61 0.55 0.38 0.61 0.0039 0.9973 注：Std、MAE、MSE、RMSE、Bias数据仅保留2位小数。
下载: 导出CSV
-
[1] Martin S. 海洋遥感导论[M]. 蒋兴伟, 译. 北京: 海洋出版社, 2008.Martin S. An Introduction to Ocean Remote Sensing[M]. Jiang Xingwei, trans. Beijing: Ocean Press, 2008. [2] Vytla V, Baduru B, Kolukula S S, et al. Forecasting of sea surface temperature using machine learning and its applications[J]. Journal of Earth System Science, 2025, 134(1): 25. doi: 10.1007/s12040-024-02483-0 [3] Lau N C, Nath M J. Impact of ENSO on SST Variability in the North Pacific and North Atlantic: seasonal dependence and role of extratropical sea–air coupling[J]. Journal of Climate, 2001, 14(13): 2846−2866. doi: 10.1175/1520-0442(2001)014<2846:IOEOSV>2.0.CO;2 [4] Mellin C, Brown S, Heron S F, et al. CoralBleachRisk—global projections of coral bleaching risk in the 21st century[J]. Global Ecology and Biogeography, 2025, 34(2): e13955. doi: 10.1111/geb.13955 [5] Oliver E C J, Benthuysen J A, Bindoff N L, et al. The unprecedented 2015/16 Tasman Sea marine heatwave[J]. Nature Communications, 2017, 8(1): 16101. doi: 10.1038/ncomms16101 [6] Chen T C, Kahru M, Landry M R, et al. Multi-trophic level responses to marine heatwave disturbances in the California current ecosystem[J]. Ecology Letters, 2024, 27(12): e14502. doi: 10.1111/ele.14502 [7] Iglesias I S, Fiechter J, Santora J A, et al. Vertical distribution of mesopelagic fishes deepens during marine heatwave in the California current[J]. ICES Journal of Marine Science, 2024, 81(9): 1837−1849. doi: 10.1093/icesjms/fsae129 [8] 胡秋良. 渤黄海温、流季节变异研究及对海上军事活动的影响[D]. 青岛: 中国海洋大学, 2005. Doi: 10.7666/d.y828432.Hu Qiuliang. Temperature and current seasonal variation of the Bohai Sea and Huanghai Sea and its effect of operation on sea[D]. Qingdao: Ocean University of China. Doi: 10.7666/d.y828432. [9] Roemmich D, Johnson G C, Riser S, et al. The Argo program: observing the global ocean with profiling floats[J]. Oceanography, 2009, 22(2): 34−43. doi: 10.5670/oceanog.2009.36 [10] Reynolds R W, Rayner N A, Smith T M, et al. An improved in situ and satellite SST analysis for climate[J]. Journal of Climate, 2002, 15(13): 1609−1625. doi: 10.1175/1520-0442(2002)015<1609:AIISAS>2.0.CO;2 [11] Guan Lei, Kawamura H. SST availabilities of satellite infrared and microwave measurements[J]. Journal of Oceanography, 2003, 59(2): 201−209. doi: 10.1023/A:1025543305658 [12] Dommenget D. An objective analysis of the observed spatial structure of the tropical Indian Ocean SST variability[J]. Climate Dynamics, 2011, 36(11/12): 2129−2145. doi: 10.1007/s00382-010-0787-1 [13] 张大明, 许东峰, 章本照, 等. 最优插值法及其在热带太平洋海表温度数据同化中的应用[J]. 海洋学研究, 2005, 23(4): 1−7. doi: 10.3969/j.issn.1001-909X.2005.04.001Zhang Daming, Xu Dongfeng, Zhang Benzhao, et al. Optimal interpolation and its application to assimilation of SST data in the Tropic Pacific Ocean[J]. Journal of Marine Sciences, 2005, 23(4): 1−7. doi: 10.3969/j.issn.1001-909X.2005.04.001 [14] Waters J, Lea D J, Martin M J, et al. Implementing a variational data assimilation system in an operational 1/4 degree global ocean model[J]. Quarterly Journal of the Royal Meteorological Society, 2015, 141(687): 333−349. doi: 10.1002/qj.2388 [15] Guo Peng. Study on Bayesian hierarchal model-based SST data fusion methods[C]//Remote Sensing of the Ocean, Sea Ice, and Large Water Regions 2010. Toulouse, France: SPIE, 2010, 7825: 78250O. Doi: 10.1117/12.864912. [16] Dong Wanqiu, Han Guijun, Li Wei, et al. A comparative evaluation of two bias correction approaches for SST forecasting: data assimilation versus deep learning strategies[J]. Remote Sensing, 2025, 17(9): 1602. doi: 10.3390/rs17091602 [17] Ma Yuanzhe, Xie Bowen, Feng Zhongkun, et al. A deep learning-based hybrid model for improved SST prediction in the tropical Pacific Ocean[J]. Journal of Oceanology and Limnology, 2025, 43(6): 1709−1725. doi: 10.1007/s00343-025-4333-8 [18] 袁本坤, 黄蕊, 商杰, 等. 基于岸基观测数据的渤海沿岸海域表层温盐特征分析[J]. 海洋开发与管理, 2015, 32(12): 31−34.Yuan Benkun, Huang Rui, Shang Jie, et al. Analysis of sea surface temperature and salinity based on coastal observed data in the Bohai Sea[J]. Ocean Development and Management, 2015, 32(12): 31−34. [19] 鲍献文, 万修全, 高郭平, 等. 渤海、黄海、东海AVHRR海表温度场的季节变化特征[J]. 海洋学报, 2002, 24(5): 125−133. doi: 10.3321/j.issn:0253-4193.2002.05.015Bao Xianwen, Wan Xiuquan, Gao Guoping, et al. The characteristics of the seasonal variability of the sea surface temperature field in the Bohai Sea, the Huanghai Sea and the East China Sea from AVHRR data[J]. Haiyang Xuebao, 2002, 24(5): 125−133. doi: 10.3321/j.issn:0253-4193.2002.05.015 [20] Qin Huiling, Chen Guixing, Wang Weiqiang, et al. Validation and application of MODIS-derived SST in the South China Sea[J]. International Journal of Remote Sensing, 2014, 35(11/12): 4315−4328. doi: 10.1080/01431161.2014.916439 [21] Gentemann C L, Akella S. Evaluation of NASA GEOS-ADAS Modeled Diurnal Warming Through Comparisons to SEVIRI and AMSR2 SST Observations[J]. Journal of Geophysical Research: Oceans, 2018, 123(2): 1364−1375. doi: 10.1002/2017JC013186 [22] Jung S, Yoo C, Im J. High-resolution seamless daily sea surface temperature based on satellite data fusion and machine learning over Kuroshio Extension[J]. Remote Sensing, 2022, 14(3): 575. doi: 10.3390/rs14030575 [23] Shi Xinjie, Duan Boheng, Ren Kaijun. A more accurate field-to-field method towards the wind retrieval of HY-2B scatterometer[J]. Remote Sensing, 2021, 13(12): 2419. doi: 10.3390/rs13122419 [24] Cornford D, Nabney I T, Ramage G. Improved neural network scatterometer forward models[J]. Journal of Geophysical Research: Oceans, 2001, 106(C10): 22331−22338. doi: 10.1029/2000JC000417 [25] Harbola S, Coors V. One dimensional convolutional neural network architectures for wind prediction. Energy Conversion and Management, 2019, 195: 70−75. [26] Kattenborn T, Leitloff J, Schiefer F, et al. Review on Convolutional Neural Networks (CNN) in vegetation remote sensing. ISPRS Journal of Photogrammetry and Remote Sensing, 2021, 173: 24−49. [27] Breiman L. Random forests. Machine Learning, 2001, 45(1): 5−32. [28] Breiman L. Bagging predictors[J]. Machine Learning, 1996, 24(2): 123−140. doi: 10.1023/A:1018054314350 -
计量
- 文章访问数: 105
- HTML全文浏览量: 52
- PDF下载量: 12
- 被引次数: 0
