基于机器学习的海洋浮标寿命及轨迹预测

Argo计划(Array or Real-time Geostrophic Oceanography)为海洋和大气研究提供了宝贵的资料,在短期天气预报和长期气候预测中起到了重要作用.为保证Argo观测阵列的正常运转,需要时刻关注浮标的运行状态,以保证研究区域内维持一定数量和密度的浮标.然而Argo浮标投放费用高昂,投放...     

Common-offset domain reflection–diffraction separation with deep learning

Separation of diffracted from reflected events in seismic data is still challenging due to the relatively low amplitude of the diffracted wavefield compared to the reflected wavefield as well as the overlap in the kinematics of reflection and diffraction events. A workflow based on deep learning can be a simple and fast alternative, but using training data made by physics-based modelling is expensive and lacks diversity in terms of noise, amplitude, frequency content and wavelet. This results in poor generalization beyond the training data without retraining and transfer learning. In this paper, we demonstrate successful applications of reflection–diffraction separation using a conventional U-net architecture. The novelty of our approach is that we do not use synthetic data created from physics-based modelling, but instead use only synthetic data built from basic geometric shapes. Our domain of application is the pre-migration common-offset domain where reflected events resemble local geology and the diffracted wavefield consists of downward convex hyperbolic diffraction patterns. Both patterns were randomly perturbed in many ways while maintaining their intrinsic features. This approach is inspired by the common practice of data augmentation in deep learning for machine vision applications. Since many of the standard data augmentation techniques lack a geophysical motivation, we have instead perturbed our synthetic training data in ways to make more sense from a signal processing perspective or given our ‘domain knowledge’ of the problem at hand. We did not have to retrain the network to show good results on the field data set. The large variety and diversity in examples enabled to trained neural networks to show encouraging results on synthetic and field data sets that were not used in training.     

Machine learning-based automatic construction of earthquake catalog for reservoir areas in multiple river basins of Guizhou province,China

Large reservoirs have the risk of reservoir induced seismicity. Accurately detecting and locating microseismic events are crucial when studying reservoir earthquakes. Automatic earthquake monitoring in reservoir areas is one of the effective measures for earthquake disaster prevention and mitigation. In this study, we first applied the automatic location workflow (named LOC-FLOW) to process 14-day continuous waveform data from several reservoir areas in different river basins of Guizhou province. Compared with the manual seismic catalog, the recall rate of seismic event detection using the workflow was 83.9%. Of the detected earthquakes, 88.9% had an onset time difference below 1 s, 81.8% has a deviation in epicenter location within 5 km, and 77.8% had a focal depth difference of less than 5 km, indicating that the workflow has good generalization capacity in reservoir areas. We further applied the workflow to retrospectively process continuous waveform data recorded from 2020 to the first half of 2021 in reservoir areas in multiple river basins of western Guizhou province and identified five times the number of seismic events obtained through manual processing. Compared with manual processing of seismic catalog, the completeness magnitude had decreased from 1.3 to 0.8, and a b-value of 1.25 was calculated for seismicity in western Guizhou province, consistent with the b-values obtained for the reservoir area in previous studies. Our results show that seismicity levels were relatively low around large reservoirs that were impounded over 15 years ago, and there is no significant correlation between the seismicity in these areas and reservoir impoundment. Seismicity patterns were notably different around two large reservoirs that were only impounded about 12 years ago, which may be explained by differences in reservoir storage capacity, the geologic and tectonic settings, hydrogeological characteristics, and active fault the reservoir areas. Prominent seismicity persisted around two large reservoirs that have been impounded for less than 10 years. These events were clustered and had relatively shallow focal depths. The impoundment of the Jiayan Reservoir had not officially begun during this study period, but earthquake location results suggested a high seismicity level in this reservoir area. Therefore, any seismicity in this reservoir area after the official impoundment deserves special attention.     

Benchmark on the accuracy and efficiency of several neural network based phase pickers using datasets from China Seismic Network

Seismic phase pickers based on deep neural networks have been extensively used recently, demonstrating their advantages on both performance and efficiency. However, these pickers are trained with and applied to different data. A comprehensive benchmark based on a single dataset is therefore lacking. Here, using the recently released DiTing dataset, we analyzed performances of seven phase pickers with different network structures, the efficiencies are also evaluated using both CPU and GPU devices. Evaluations based on F₁-scores reveal that the recurrent neural network (RNN) and EQTransformer exhibit the best performance, likely owing to their large receptive fields. Similar performances are observed among PhaseNet (UNet), UNet++, and the lightweight phase picking network (LPPN). However, the LPPN models are the most efficient. The RNN and EQTransformer have similar speeds, which are slower than those of the LPPN and PhaseNet. UNet++ requires the most computational effort among the pickers. As all of the pickers perform well after being trained with a large-scale dataset, users may choose the one suitable for their applications. For beginners, we provide a tutorial on training and validating the pickers using the DiTing dataset. We also provide two sets of models trained using datasets with both 50 Hz and 100 Hz sampling rates for direct application by end-users. All of our models are open-source and publicly accessible.     

基于集成学习的海岸带变化检测方法研究

随着我国自主研发卫星组网的不断完善,利用遥感变化检测技术进行海岸带变化检测成为海岸带监测的重要手段。针对沿海地区的变化信息提取,文章首先利用多特征构建差异影像,在此基础上采用两种集成学习方式:随机森林（Random Forest）和极端梯度提升（XGBoost）,进行试验区的变化检测,并与传统的机器学习SVM、经典的变化检测方法CVA和IR-MAD进行对比,结果表明集成学习进行变化信息提取效率远超其余方式,且XGBoost在变化信息提取精度方面具有一定优势。研究成果对海岸带及海域使用开展自动化变化监测和海岸带监督管理具有重要意义。     

A comparative study of three different learning algorithms applied to ANFIS for predicting daily suspended sediment concentration

The modeling and prediction of suspended sediment in a river are key elements in global water recourses and environment policy and management. In the present study, an Adaptive Neuro-Fuzzy Inference System model trained with the Levenberg-Marquardt learning algorithm is considered for time series modeling of suspended sediment concentration in a river. The model is trained and validated using daily river discharge and suspended sediment concentration data from the Schuylkill River in the United States. The results of the proposed method are evaluated and compared with similar networks trained with the common Hybrid and Back-Propagation algorithms, which are widely used in the literature for prediction of suspended sediment concentration. Obtained results demonstrate that models trained with the Hybrid and Levenberg-Marquardt algorithms are comparable in terms of prediction accuracy. However, the networks trained with the Levenberg-Marquardt algorithm perform better than those trained with the Hybrid approach.     

基于压缩感知的多跳地震数据采集技术与方法

随着油气地震勘探目标的复杂程度日益提高,地震数据采集系统的道容量也需要得到进一步的提升.本文根据压缩感知、稀疏表示等理论,提出了一种多跳恒传输量的数据采集框架,以减少每条测线上地震数据的传输量,进而提升采集系统的道容量.为了能够明显地提高带道能力,设计了基于有序并行原子更新的字典学习算法,该算法能够在计算量较小的前提下有效的得到相应数据的稀疏变换矩阵.基于压缩感知的多跳地震数据采集方法已能够在吉林大学研制的无缆自定位地震仪中实现.本文最后使用一组仿真数据和一组实测数据进行测试,其结果表明,数据采集信噪比控制在14 dB以上(引入噪声约18%)时,最多可以将系统的道容量提高3倍以上.     

Assessment and evaluation of potential climate change impact on monsoon flows using machine learning technique over Wainganga River basin,India

In this study, classification- and regression-based statistical downscaling is used to project the monthly monsoon streamflow over the Wainganga basin, India, using 40 global climate model (GCM) outputs and four representative concentration pathways (RCP) scenarios. Support vector machine (SVM) and relevance vector machine (RVM) are considered to perform downscaling. The RVM outperforms SVM and is used to simulate future projections of monsoon flows for different periods. In addition, variability in water availability with uncertainty and change point (CP) detection are accomplished by flow–duration curve and Bayesian analysis, respectively. It is observed from the results that the upper extremes of monsoon flows are highly sensitive to increases in temperature and show a continuous decreasing trend. Medium and low flows are increasing in future projections for all the scenarios, and high uncertainty is noticed in the case of low flows. An early CP is detected in the case of high emissions scenarios.     

基于BP-Adaboost方法的天然地震和人工爆炸事件波形信号分类识别研究

BP神经网络和支持向量机(SVM)是两种主流的分类识别方法,用于天然地震和人工爆炸事件波形信号分类识别时取得了较好的效果。但BP神经网络存在易陷入局部最优及隐层数和隐层节点数与训练样本数据密切相关而无法有效预先确定;而支持向量机(SVM)方法则缺乏有效手段来选取合适的核函数,从中不能很好地扩展到多分类。针对天然地震和人工爆炸事件波形信号的分类识别问题,文中将上述两种方法和集成学习——BP-Adaboost方法进行了对比实验研究。据对所选用的地震、爆炸事件波形信号数据集的分类识别结果表明,BP-Adaboost方法得到了98%以上的正确识别率,并且具有较好的泛化能力。相较于BP神经网络和PCA-SVM方法,BP-Adaboost方法对于数据集的划分和识别结果具有更好的鲁棒性,应用于天然地震和人工爆炸事件波形信号分类识别时,可取得更好的识别效果。同时,结合Adaboost方法的原理,阐述了BP-Adaboost方法拥有更好分类结果和泛化能力的原因。     

Deep learning for dune pattern mapping with the AW3D30 global surface model

In this paper we present a deep learning (U-Net)-based workflow for classifying linear dune landforms based on the discrete Laplacian convolution of a new global elevation dataset, the AW3D30 digital surface model. Crest vectors were then derived for landscape pattern analysis. The U-Net crest classification model was trained and evaluated on sample data from dunefields across the Australian continent. The resulting crest vectors and dune defect placement were then evaluated in typical semi-arid and arid dune landscapes in eastern central Australia where high-resolution (5 m horizontal) digital elevation models are available (for three out of our four study sites) as a reference dataset. The method was applied to quantify dune pattern metrics for the entire Simpson Desert dunefield, Australia. The U-Net does a very good job of segmenting dune crests, even where dunes are less clear in the Laplacian map (intersection over union score ≈ 0.68). When crest vectors and dune defects (network nodes) were derived, the defect predictions were typically correct (0.4 to 0.79 correctness) but incomplete (0.02 to 0.64 completeness). Much of the residual error was traced to the resolution of the input data. Through the application to the Simpson Desert, we nevertheless demonstrated that our method can effectively be used for regional-scale dune pattern analysis. Furthermore, we suggest that the combination of morphological filtering and a convolutional neural network could readily be adapted to target other geomorphic features, such as channel networks or geological lineaments. © 2020 John Wiley & Sons, Ltd.