A spatiotemporal prediction method for Arctic monthly mean sea ice concentration based on an improved SA-ConvLSTM model
-
摘要: 针对融冰期北极海冰密集度预测精度不高的问题，本文构建了一种基于改进的SA-Conv-LSTM模型进行北极海冰密集度预测的方法，用于实现未来1年内月均海冰密集度数据的二维时空预测。该方法以SA-ConvLSTM模型为核心单元，通过引入Seq2Seq预测结构和类VGG16编解码器结构，有针对性地解决时间序列输出步长选择过程不确定的问题，并设置一种组合损失函数来优化训练过程，以进一步提升海冰密集度分布的时空预测精度。以北冰洋为实验区，基于美国国家冰雪数据中心（NSIDC）与国家海洋和大气管理局（NOAA）联合发布的海冰密集度气候月均数据，预测了2023年北极海冰密集度的时空分布，并与真实数据进行了对比分析。结果表明：与传统LSTM、ConvLSTM以及未改进的SA-ConvLSTM模型相比，本文改进模型在各项指标上均表现出较大优势，其中：均方根误差分别下降13.18%、36.10%和22.58%；相关系数分别提高1.90%、5.97%和3.31%；结构相似性指数分别增加5.38%、15.00%和10.30%；海冰面积偏差分别降低了83.46%、76.53%和60.30%。此外，通过对2012年与2020年极端融冰年份的预测结果分析，进一步验证了该模型在异常气候条件下的稳定性与鲁棒性，显示出良好的适应性与实际应用潜力。本文时空预测模型在融冰期能够更准确地预测海冰的空间分布，能捕捉复杂的时空变化信息及细节。
-
关键词:
- 时空预测 /
- 海冰密集度 /
- SA-ConvLSTM /
- 组合损失函数 /
- 自注意力
Abstract: To address the problem of low prediction accuracy for Arctic sea ice concentration during the melting season, this study proposes a method for predicting Arctic sea ice concentration based on an improved SA-ConvLSTM model, enabling two-dimensional spatiotemporal prediction of monthly mean sea ice concentration data for the coming year. The method uses the SA-ConvLSTM model as the core unit, incorporating a Seq2Seq prediction structure and a VGG16-like encoder-decoder architecture to specifically address the uncertainty in selecting the output step length of time series predictions. In addition, a composite loss function is designed to optimize the training process, further enhancing the spatiotemporal prediction accuracy of sea ice concentration distribution. Using the Arctic Ocean as the study area, and based on monthly climate sea ice concentration data jointly released by the National Snow and Ice Data Center (NSIDC) and the National Oceanic and Atmospheric Administration (NOAA), the model predicts the spatiotemporal distribution of Arctic sea ice concentration in 2023 and compares the results with real observations. The results show that, compared with traditional LSTM, ConvLSTM, and the unmodified SA-ConvLSTM models, the improved model achieves significant advantages in all evaluation metrics: root mean square error decreases by 13.18%, 36.10%, and 22.58%; correlation coefficient increases by 1.90%, 5.97%, and 3.31%; structural similarity index improves by 5.38%, 15.00%, and 10.30%; and sea ice area error decreases by 83.46%, 76.53%, and 60.30%, respectively. Furthermore, analysis of predictions for the extreme melting years of 2012 and 2020 further verifies the model’s stability and robustness under abnormal climatic conditions, demonstrating strong adaptability and practical application potential. The proposed spatiotemporal prediction model can more accurately predict the spatial distribution of sea ice during the melting season and effectively capture complex spatiotemporal variations and fine-scale details. -
图 7 本文模型2023年预测结果及真实海冰结果
Fig. 7 The model prediction results in 2023 and real sea ice results in this paper
图 9 ConvLSTM模型2023年海冰预测结果
Fig. 9 The sea ice prediction results of the ConvLSTM model in 2023
图 10 未引入组合损失函数的SA-ConvLSTM模型2023年海冰预测结果
Fig. 10 The sea ice prediction results of the SA-ConvLSTM model without introducing the combined loss function in 2023
图 12 不同预测模型结果的细节对比
Fig. 12 Detailed comparison of the results of different prediction models
图 13 北极融化季节的海冰边缘比较
Fig. 13 Comparison of sea ice margins during the Arctic melt season
图 14 2012年5−10月不同方法的预测结果及真实海冰结果对比
Fig. 14 The comparison of predicted sea ice results using different methods and real sea ice results from May to October 2012
图 15 2020年5−10月不同方法的预测结果及真实海冰结果对比
Fig. 15 The comparison of predicted sea ice results using different methods and real sea ice results from May to October 2020
表 1 实验平台主要参数信息
Tab. 1 Experimental platform main parameter information
配置 参数 处理器 12th Gen Intel(R) Core(TM) i9−12900H 2.50 GHz 内存 16.0 GB 图形处理器 NVIDIA GeForce RTX 3060 Laptop GPU 显存 6.0 GB 操作系统 Windows 11 64 bit 编程语言 Python 3.9 深度学习框架 Pytorch 1.13 
下载: 导出CSV
表 2 不同时间步长对网络精度的影响
Tab. 2 Influence of different time steps onnetwork accuracy
时间步长/月 RMSE/
106 km2MSE/
106 km2MAE/
106 km2R2 SSIM 3 0.1640 0.0287 0.0576 0.7528 0.7455 6 0.1131 0.0133 0.0343 0.7778 0.8340 12 0.0924 0.0086 0.0279 0.8768 0.8483 18 0.0996 0.0098 0.0336 0.8643 0.8045
下载: 导出CSV
表 3 不同模型的年均预测精度
Tab. 3 Annual average prediction accuracy ofdifferent models
年份 模型 RMSE/106 km2 R2 SSIM SIAE/106 km2 2012 LSTM 0.0847 0.8753 0.8919 1.4057 ConvLSTM 0.1050 0.8508 0.8072 0.9200 本文方法 0.0493 0.9603 0.9444 0.4642 2020 LSTM 0.0735 0.9021 0.8898 1.8183 ConvLSTM 0.1086 0.8310 0.8079 1.0558 本文方法 0.0482 0.9582 0.9487 0.4194 2023 LSTM 0.0683 0.9239 0.8934 1.4144 ConvLSTM 0.0928 0.8884 0.8187 1.0433 本文方法 0.0593 0.9415 0.9415 0.2448
下载: 导出CSV
-
[1] Sumata H, de Steur L, Divine D V, et al. Regime shift in Arctic Ocean sea ice thickness[J]. Nature, 2023, 615(7952): 443−449. doi: 10.1038/s41586-022-05686-x [2] Ingvaldsen R B, Assmann K M, Primicerio R, et al. Physical manifestations and ecological implications of Arctic Atlantification[J]. Nature Reviews Earth & Environment, 2021, 2(12): 874−889. [3] 田康, 王志勇, 李振今, 等. 格陵兰彼得曼冰川表面运动特征和季节年度变化分析[J]. 测绘通报, 2024(7): 35−40,54.Tian Kang, Wang Zhiyong, Li Zhenjin, et al. Analysis of surface movement characteristics and seasonal annual variation of Petermann Glacier in Greenland[J]. Bulletin of Surveying and Mapping, 2024(7): 35−40,54. [4] 曹云锋, 于萌, 惠凤鸣, 等. 北极冰区通航能力变化研究进展[J]. 科学通报, 2021, 66(1): 21−33. doi: 10.1360/TB-2020-0596Cao Yunfeng, Yu Meng, Hui Fengming, et al. Review of navigability changes in trans-Arctic routes[J]. Chinese Science Bulletin, 2021, 66(1): 21−33. doi: 10.1360/TB-2020-0596 [5] Sardain A, Sardain E, Leung B. Global forecasts of shipping traffic and biological invasions to 2050[J]. Nature Sustainability, 2019, 2(4): 274−282. doi: 10.1038/s41893-019-0245-y [6] 刘泉宏, 张韧, 汪杨骏, 等. 深度学习方法在北极海冰预报中的应用[J]. 大气科学学报, 2022, 45(1): 14−21.Liu Quanhong, Zhang Ren, Wang Yangjun, et al. Application of deep learning methods to Arctic sea ice prediction[J]. Transactions of Atmospheric Sciences, 2022, 45(1): 14−21. [7] Horvath S, Stroeve J, Rajagopalan B, et al. A Bayesian logistic regression for probabilistic forecasts of the minimum September Arctic sea ice cover[J]. Earth and Space Science, 2020, 7(10): e2020EA001176. doi: 10.1029/2020EA001176 [8] Reichstein M, Camps-Valls G, Stevens B, et al. Deep learning and process understanding for data-driven Earth system science[J]. Nature, 2019, 566(7743): 195−204. doi: 10.1038/s41586-019-0912-1 [9] Chi J, Kim H C. Prediction of Arctic sea ice concentration using a fully data driven deep neural network[J]. Remote Sensing, 2017, 9(12): 1305. doi: 10.3390/rs9121305 [10] Kim Y J, Kim H C, Han D, et al. Prediction of monthly Arctic sea ice concentrations using satellite and reanalysis data based on convolutional neural networks[J]. The Cryosphere, 2020, 14(3): 1083−1104. doi: 10.5194/tc-14-1083-2020 [11] Christofi V P, Wang Xiaodi. Optimizing the LSTM deep learning model for Arctic sea ice melting prediction[J]. Atmospheric and Climate Sciences, 2024, 14(4): 429−449. doi: 10.4236/acs.2024.144026 [12] Moishin M, Deo R C, Prasad R, et al. Designing deep-based learning flood forecast model with ConvLSTM hybrid algorithm[J]. IEEE Access, 2021, 9: 50982−50993. doi: 10.1109/ACCESS.2021.3065939 [13] Liu Quanhong, Zhang Ren, Wang Yangjun, et al. Daily prediction of the Arctic sea ice concentration using reanalysis data based on a convolutional LSTM network[J]. Journal of Marine Science and Engineering, 2021, 9(3): 330. doi: 10.3390/jmse9030330 [14] Li Penglun, Leng Hongze, Ma Beibei, et al. Monthly Arctic sea ice concentration prediction based on the ConvLSTM model[C]//IEEE 2021 International Conference on Computer Information Science and Artificial Intelligence (CISAI). Kunming: IEEE, 2021: 162−167. [15] Andersson T R, Hosking J S, Pérez-Ortiz M, et al. Seasonal Arctic sea ice forecasting with probabilistic deep learning[J]. Nature Communications, 2021, 12(1): 5124. doi: 10.1038/s41467-021-25257-4 [16] Ren Yibin, Li Xiaofeng, Zhang Wenhao. A data-driven deep learning model for weekly sea ice concentration prediction of the pan-Arctic during the melting season[J]. IEEE Transactions on Geoscience and Remote Sensing, 2022, 60: 4304819. [17] Fu En, Zhang Yinong, Yang Fan, et al. Temporal self-attention-based Conv-LSTM network for multivariate time series prediction[J]. Neurocomputing, 2022, 501: 162−173. doi: 10.1016/j.neucom.2022.06.014 [18] Zheng Qingyu, Wang Ru, Han Guijun, et al. A spatiotemporal multiscale deep learning model for subseasonal prediction of Arctic sea ice[J]. IEEE Transactions on Geoscience and Remote Sensing, 2024, 62: 4300522. [19] 孙士昌, 王志勇, 李振今, 等. 基于改进DeepLabV3+模型的海冰提取方法——以北极格陵兰海为例[J]. 海洋学报, 2024, 46(8): 131−142.Sun Shichang, Wang Zhiyong, Li Zhenjin, et al. An extraction method for sea ice based on improved DeepLabV3+ model: taking the Arctic Greenland Sea as an example[J]. Haiyang Xuebao, 2024, 46(8): 131−142. [20] Ge Hongxia, Li Siteng, Cheng Rongjun, et al. Self-attention ConvLSTM for spatiotemporal forecasting of short-term online car-hailing demand[J]. Sustainability, 2022, 14(12): 7371. doi: 10.3390/su14127371 [21] Fang Wei, Chen Yupeng, Xue Qiongying. Survey on research of RNN-based spatio-temporal sequence prediction algorithms[J]. Journal on Big Data, 2021, 3(3): 97−110. doi: 10.32604/jbd.2021.016993 [22] Sriram A, Jun H, Satheesh S, et al. Cold fusion: training Seq2Seq models together with language models[C]//19th Annual Conference of the International Speech Communication Association. Hyderabad: ISCA, 2018: 387−391. [23] Chen Mingjun, Bovik A C. Fast structural similarity index algorithm[J]. Journal of Real-Time Image Processing, 2011, 6(4): 281−287. doi: 10.1007/s11554-010-0170-9 [24] 李朋伦. 基于深度学习的全球海冰资料模拟预测研究[D]. 长沙: 国防科技大学, 2021.Li Penglun. Simulation and prediction of global sea ice data based on deep learning[D]. Changsha: National University of Defense Technology, 2021. -
计量
- 文章访问数: 184
- HTML全文浏览量: 144
- PDF下载量: 21
- 被引次数: 0
