Multiscale attentional coding-dynamic decoding network for short-range precipitation nowcasting
-
摘要: 短临降水预报是气象学和水文学中的重要任务之一，但在现有深度学习方法中，其预测结果模糊不清，并且累计误差大。为了克服这些预测方法中存在累计误差的局限性，以及预测序列结果模糊不清的问题，本文构建了一种基于多尺度注意力编码−动态解码网络（Multi-scale Attention Encoding-Dynamic Decoding Network, MAEDDN）的短临降水预报方法，通过学习输入数据的时空特征来预测未来的降水情况。为了得到更多输入序列的特征信息，在编码过程中，使用带有空间及通道注意力的卷积块进行编码，并增加多尺度融合模块解决降水分布中小尺度与大尺度信息无法同时捕获的问题；增强预测序列的清晰度，需要模型更好地理解降水过程，因此在解码过程中，针对短临降水过程伴随的生成与消散过程，提出了一种动态解码网络，通过学习输入过去数据的强度分布及变化趋势对解码过程进行灵活地筛选。使用公开数据集SEVIR的降水数据进行实验，并与现有最好模型进行对比，实验结果表明：（1）MAEDDN提升了在高强度降水区域的预测能力；（2）MAEDDN预测的图像序列清晰度显著优于其他模型。构建的多尺度注意力编码能够更好地捕捉气象数据的复杂关系；动态解码能够根据不同的情况自适应地选择解码过程，提供更准确的预测结果。Abstract: Short-term precipitation nowcasting is a critical task in both meteorology and hydrology. However, current deep learning methods often yield ambiguous prediction results and exhibit significant cumulative errors. To address the limitations associated with these predictive methods, particularly the challenges of cumulative error and lack of clarity in prediction sequences, we propose a novel approach based on a Multi-scale Attention Encoding-Dynamic Decoding Network (MAEDDN) for short-term precipitation nowcasting. This method leverages the learning of spatiotemporal features from input data to accurately predict future precipitation scenarios. To obtain richer feature information from the input sequences, the encoding process employs convolutional blocks with spatial and channel attention for encoding. And a multi-scale fusion module is introduced to address the challenge of capturing both small-scale and large-scale information in precipitation distribution simultaneously. To enhance the clarity of the predicted sequences, the model needs to better understand the precipitation process. Therefore, in the decoding process, a dynamic decoding network is proposed in response to the generation and dissipation processes accompanying short-term precipitation. This network flexibly filters the decoding process by learning the intensity distribution and change trends of past input data. Experiments are conducted by using the precipitation data from the open-source SEVIR dataset, and comparisons are made with the best methods reported so far. The experimental results reveal that: (1) MAEDDN enhances the forecasting capability in areas with high-intensity precipitation, and (2) The clarity of the predicted image sequences by MAEDDN is significantly better than that of other models. The constructed multi-scale attention encoding captures the complex relationships in meteorological data more effectively, while the dynamic decoding adapts the decoding process based on different scenarios, resulting in more accurate prediction outcomes.
-
图 1 多尺度注意力编码−动态解码网络结构
a. 多尺度注意力编码模块，b. 动态路由解码模块
Fig. 1 Multi-scale attention encoding-dynamic decoding network structure
a. Multi-scale attention encoding module, b. dynamic routing decoding module
图 2 多尺度特征注意力模块(MAMB)架构的详细信息
Fig. 2 Detailed information on the Multi-scale Attention Feature Module (MAMB) architecture
图 4 反射率阈值在16（a）、133（b）、219（c）的10个时间步外推过程中的指标变化
Fig. 4 Variation of metrics during the extrapolation process over 10 time steps with reflectivity thresholds of 16 (a), 133 (b) and 219 (c)
图 5 对比实验预测图像序列结果（生成）
Fig. 5 Results of the generated image sequences from the comparison experiments (generation)
图 6 对比实验预测图像序列结果（消散）
Fig. 6 Results of the generated image sequences from the comparison experiments (dissipation)
图 7 消融实验预测图像序列结果
Fig. 7 Results of the generated image sequences from the ablation experiments
表 1 MAMB中卷积神经网络参数配置
Tab. 1 Parameter configuration for convolutional neural networks in MAMB
F1主干道卷积神经网络 类型 卷积核 步长 Conv 3 × 3 × 150 1 × 1 PReLU Conv 3 × 3 × 300 1 × 1 PReLU Conv 3 × 3 × 300 1 × 1 PReLU Conv 3 × 3 × 300 1 × 1 PReLU Conv 3 × 3 × 75 1 × 1 PReLU F2副干道卷积神经网络 类型 卷积核 步长 Conv 3 × 3 × 32 2 × 2 PReLU Conv 3 × 3 × 32 1 × 1 PReLU 
下载: 导出CSV
表 2 DDRB中卷积神经网络参数配置
Tab. 2 Parameter configuration for convolutional neural networks in DDRB
类型 卷积核 步长 Conv 3 × 3 × 32 1 × 1 ReLU Conv 3 × 3 × 32 1 × 1 Conv 3 × 3 × 32 1 × 1 Conv 3 × 3 × 4 1 × 1
下载: 导出CSV
表 3 实验环境配置
Tab. 3 Experimental environment configuration
名称 相关配置 操作系统 Linux-5.15.0 处理器 Intel Core i7-12700K 内存 64 GB 显卡 NVIDIA 3060 深度学习框架 PyTorch 1.8.1
下载: 导出CSV
表 4 数据集详细配置
Tab. 4 Detailed configuration of the dataset
数据集 事件数 图像数 模式 SEVIR_VIL_2017 2778 55560 训练 SEVIR_VIL_2018 1712 34240 测试
下载: 导出CSV
表 5 SEVIR数据集上与SOTA模型对比实验结果（加粗数值为该指标性能最好）
Tab. 5 Comparison of experimental results with SOTA models on the SEVIR dataset (bold values indicate the best performance for that metric)
模型 评价指标 CSI-M↑ CSI-219↑ CSI-181↑ CSI-160↑ CSI-133↑ CSI-74↑ CSI-16↑ MSE(10−3)↓ SmaAt-UNet 0.3781 0.1077 0.1480 0.2557 0.3404 0.6731 0.7440 4.0015 ConvLSTM 0.4185 0.1288 0.2482 0.2928 0.4052 0.6793 0.7569 3.7532 PredRNN 0.4080 0.1312 0.2324 0.2767 0.3858 0.6713 0.7507 3.9014 PhyDnet 0.3940 0.1288 0.2309 0.2708 0.3720 0.6556 0.7059 4.8165 BEP 0.3843 0.1105 0.1629 0.2650 0.3506 0.6730 0.7439 3.9868 MAEDDN(ours) 0.4523 0.2330 0.3077 0.3645 0.3945 0.6514 0.7629 3.4463
下载: 导出CSV
表 6 提出网络的消融实验结果（CSI）（加粗数值为该指标性能最好）
Tab. 6 Ablation experimental results of the proposed network (CSI) (bold values indicate the best performance for that metric)
模型设置 评价指标 CSI-M↑ CSI-219↑ CSI-181↑ CSI-160↑ CSI-133↑ CSI-74↑ CSI-16↑ Backbone 0.3563 0.0705 0.2295 0.2587 0.2741 0.6047 0.7006 +MAM 0.3992 0.1501 0.2496 0.3047 0.3135 0.6274 0.7504 +MAM+DDRB 0.4523 0.2330 0.3077 0.3645 0.3945 0.6514 0.7629
下载: 导出CSV
-
[1] Chen Chengcheng, Zhang Qian, Kashani M H, et al. Forecast of rainfall distribution based on fixed sliding window long short-term memory[J]. Engineering Applications of Computational Fluid Mechanics, 2022, 16(1): 248−261. doi: 10.1080/19942060.2021.2009374 [2] Shi Xingjian, Chen Zhourong, Wang Hao, et al. Convolutional LSTM network: a machine learning approach for precipitation nowcasting[C]//Proceedings of the 28th International Conference on Neural Information Processing Systems-Volume 1. Montreal: MIT Press, 2015: 802−810. [3] Shi Xingjian, Gao Zhihan, Lausen L, et al. Deep learning for precipitation nowcasting: a benchmark and a new model[C]//Proceedings of the 31st International Conference on Neural Information Processing Systems. Long Beach: Curran Associates Inc., 2017: 5622−5632. [4] Wang Yunbo, Long Mingsheng, Wang Jianmin, et al. PredRNN: recurrent neural networks for predictive learning using spatiotemporal LSTMs[C]//Proceedings of the 31st International Conference on Neural Information Processing Systems. Long Beach: Curran Associates Inc. , 2017: 879−888. [5] Wang Yunbo, Gao Zhifeng, Long Mingsheng, et al. PredRNN++: towards a resolution of the deep-in-time dilemma in spatiotemporal predictive learning[C]//Proceedings of the 35th International Conference on Machine Learning. Stockholm: PMLR, 2018: 5123−5132. [6] Xue Fuzhao, Shi Ziji, Wei Futao, et al. Go wider instead of deeper[C]//Proceedings of the 36th AAAI Conference on Artificial Intelligence. AAAI, 2022: 8779−8787. [7] 吴卓升, 张巍, 林艳, 等. 动态概率卷积神经网络在雷达回波外推中的应用[J]. 计算机应用研究, 2021, 38(7): 2125−2129.Wu Zhuosheng, Zhang Wei, Lin Yan, et al. Application of dynamic probability convolutional neural network in radar echo extrapolation[J]. Application Research of Computers, 2021, 38(7): 2125−2129. [8] Trebing K, Staǹczyk T, Mehrkanoon S. SmaAt-UNet: precipitation nowcasting using a small attention-UNet architecture[J]. Pattern Recognition Letters, 2021, 145: 178−186. doi: 10.1016/j.patrec.2021.01.036 [9] Chen Guoxing, Wang W C. Short-term precipitation prediction for contiguous United States using deep learning[J]. Geophysical Research Letters, 2022, 49(8): e2022GL097904. doi: 10.1029/2022GL097904 [10] Qiu Yunan, Lu Zhenyu, Fang Shanpu. A short-term precipitation prediction model based on spatiotemporal convolution network and ensemble empirical mode decomposition[J]. IEEE/CAA Journal of Automatica Sinica, 2022, 9(4): 738−740. doi: 10.1109/JAS.2022.105479 [11] Ayzel G, Scheffer T, Heistermann M. RainNet v1.0: a convolutional neural network for radar-based precipitation nowcasting[J]. Geoscientific Model Development, 2020, 13(6): 2631−2644. doi: 10.5194/gmd-13-2631-2020 [12] 曹伟华, 南刚强, 陈明轩, 等. 基于深度学习的京津冀地区精细尺度降水临近预报研究[J]. 气象学报, 2022, 80(4): 546−564.Cao Weihua, Nan Gangqiang, Chen Mingxuan, et al. A study on fine scale precipitation nowcasting in Beijing-Tianjin-Hebei region based on deep learning[J]. Acta Meteorologica Sinica, 2022, 80(4): 546−564. [13] Yang Shangsahng, Yuan Huiling. A customized multi-scale deep learning framework for storm nowcasting[J]. Geophysical Research Letters, 2023, 50(13): e2023GL103979. doi: 10.1029/2023GL103979 [14] Wu Hao, Liang Yuxuan, Xiong Wei, et al. Earthfarsser: versatile spatio-temporal dynamical systems modeling in one model[C]//Proceedings of the 38th AAAI Conference on Artificial Intelligence. AAAI, 2024: 15906−15914. [15] 方巍, 庞林, 张飞鸿, 等. 对抗型长短期记忆网络的雷达回波外推算法[J]. 中国图象图形学报, 2021, 26(5): 1067−1080. doi: 10.11834/jig.200316Fang Wei, Pang Lin, Zhang Feihong, et al. Radar echo extrapolation algorithm based on adversarial long short-term memory network[J]. Journal of Image and Graphics, 2021, 26(5): 1067−1080 doi: 10.11834/jig.200316 [16] Yu Wenbin, Wang Suxun, Zhang Chengjun, et al. Integrating spatio-temporal and generative adversarial networks for enhanced nowcasting performance[J]. Remote Sensing, 2023, 15(15): 3720. doi: 10.3390/rs15153720 [17] Wang Jingnan, Wang Xiaodong, Guan Jiping, et al. TAFFNet: Time-aware adaptive feature fusion network for very short-term precipitation forecasts[J]. Geophysical Research Letters, 2023, 50(15): e2023GL104370. doi: 10.1029/2023GL104370 [18] 庄潇然, 郑玉, 王亚强, 等. 基于深度学习的融合降水临近预报方法及其在中国东部地区的应用研究[J]. 气象学报, 2023, 81(2): 286−303.Zhuang Xiaoran, Zheng Yu, Wang Yaqiang, et al. A deep learning-based precipitation nowcast model and its application over East China[J]. Acta Meteorologica Sinica, 2023, 81(2): 286−303. [19] Qiu Yunan, Lu Zhenyu, Tang Haibo. A short-term regional precipitation prediction model based on wind-improved spatiotemporal convolutional network[J]. Earth and Space Science, 2022, 9(9): e2022EA002411. doi: 10.1029/2022EA002411 [20] Fang Wei, Shen Liang, Sheng V S, et al. A novel method for precipitation nowcasting based on ST-LSTM[J]. Computers, Materials & Continua, 2022, 72(3): 4867−4877. [21] Xiong Taisong, He Jianxing, Wang Hao, et al. Contextual Sa-attention convolutional LSTM for precipitation nowcasting: a spatiotemporal sequence forecasting view[J]. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2021, 14: 12479−12491. doi: 10.1109/JSTARS.2021.3128522 [22] Wu Dali, Wu Li, Zhang Tao, et al. Short-term rainfall prediction based on radar echo using an improved self-attention PredRNN deep learning model[J]. Atmosphere, 2022, 13(12): 1963. doi: 10.3390/atmos13121963 [23] Lin Fudong, Yuan Xu, Zhang Yihe, et al. Comprehensive transformer-based model architecture for real-world storm prediction[C]//European Conference on Machine Learning and Knowledge Discovery in Databases: Applied Data Science and Demo Track. Turin: Springer, 2023: 54−71. [24] Küçük Ç, Giannakos A, Schneider S, et al. Transformer-based nowcasting of radar composites from satellite images for severe weather[J]. Artificial Intelligence for the Earth Systems, 2024, 3(2): e230041. [25] Woo S, Park J, Lee J Y, et al. CBAM: convolutional block attention module[C]//Proceedings of the 15th European Conference on Computer Vision. Munich: Springer, 2018: 3−19. [26] Kim B, Han M, Shim H, et al. A performance comparison of convolutional neural network-based image denoising methods: the effect of loss functions on low-dose CT images[J]. Medical Physics, 2019, 46(9): 3906−3923. doi: 10.1002/mp.13713 [27] Veillette M S, Samsi S, Mattioli C J. SEVIR: a storm event imagery dataset for deep learning applications in radar and satellite meteorology[C]//Proceedings of the 34th International Conference on Neural Information Processing Systems. Vancouver: Curran Associates Inc., 2020: 22009−22019. [28] Barakat A, Bianchi P. Convergence and dynamical behavior of the ADAM algorithm for nonconvex stochastic optimization[J]. SIAM Journal on Optimization, 2021, 31(1): 244−274. [29] Le Guen V, Thome N. Disentangling physical dynamics from unknown factors for unsupervised video prediction[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Seattle: IEEE, 2020: 11471−11481. [30] Yang Nan, Li Xiaofeng. Lightweight AI-powered precipitation nowcasting[J]. The Innovation Geoscience, 2024, 2(2): 100066. doi: 10.59717/j.xinn-geo.2024.100066 -
计量
- 文章访问数: 25
- HTML全文浏览量: 6
- PDF下载量: 3
- 被引次数: 0
