Volcano deformation and eruption forecasting
Volcano deformation and eruption forecasting
复制标题
火山变形和喷发预报
DOI：
--
复制
发表时间：
2013
期刊：
影响因子：
--
通讯作者：
P. Segall
中科院分区：
文献类型：
--
作者：
P. Segall
摘要 全球定位系统（GPS）、倾斜和干涉合成孔径雷达（InSAR）的最新进展极大地提高了火山形变数据的可用性。这些测量结果与适当的源模型相结合，可用于估计岩浆房深度，并提供有关岩浆房形状和体积变化的信息。然而，运动学模型无法约束岩浆室体积，并且不提供预测能力。火山爆发之前通常会经历通货膨胀时期。在适当的条件下，火山喷发是“可预测的通货膨胀”；也就是说，当通货膨胀恢复前一事件期间的通货紧缩时，随后的爆发就会发生。预测喷发方面取得的显着成功很大程度上来自于识别地震活动、地面变形和气体排放的可重复模式的能力，并结合过去喷发行为的历史和地质证据。为了超越经验模式识别，转向基于潜在动力学的确定性物理化学模型的预测，需要集成不同的数据类型和模型。我建议两个领域有望取得进展：形变和地震活动的定量整合；基于模型的预测以通过可用数据集反演对材料参数和初始条件的估计为条件。变形和地震活动是火山监测的主要地球物理方法，在某些情况下，在喷发前几分钟到几小时就发出了岩脉传播的信号。然而，缺乏与这些过程相关的定量模型。变化应力条件下地震活动率变化的现代理论可用于将变形和（火山-构造）地震活动整合到自洽反演中，以实现岩脉几何形状和过剩岩浆压力的时空演化。这种方法应该可以提高现有方法的分辨率，或许还可以改进实时预测。过去几十年里，火山喷发的物理化学模型的复杂性也显着提高。我回顾了管道模型，该模型可以通过马尔可夫链蒙特卡罗 (MCMC) 反演与 GPS 和挤压速率数据相结合，以估计地壳岩浆室的绝对体积、初始室超压、初始挥发物浓度和其他感兴趣的参数。通过使用与可用数据一致的初始条件和材料参数的分布来启动预测前向模型，MCMC 估计过程可以扩展到确定性预测。这种基于物理的 MCMC 预测将基于系统的所有知识，包括截至当前日期的数据。底层模型是完全确定性的；然而，由于该方法对与给定数据一致的初始条件和物理参数进行采样，因此它会产生包括基础参数不确定性的概率预测。因为几乎肯定会有一些影响没有被纳入前瞻模型中，因此随着模型的发展变得更加现实，可能会出现相当大的学习曲线。
Abstract Recent advances in Global Positioning System (GPS), tilt and Interferometric Synthetic Aperture Radar (InSAR) have greatly increased the availability of volcano deformation data. These measurements, combined with appropriate source models, can be used to estimate magma chamber depth, and to provide information on chamber shape and volume change. However, kinematic models cannot constrain magma chamber volume, and provide no predictive capability. Volcanic eruptions are commonly preceded by periods of inflation. Under appropriate conditions, eruptions are ‘inflation predictable’; that is, subsequent eruptions occur when inflation recovers the deflation during the preceding event. Notable successes in forecasting eruptions have come largely through the ability to discern repeatable patterns in seismic activity, ground deformation and gas emission, combined with historical and geological evidence of past eruptive behaviour. To move beyond empirical pattern recognition to forecasting based on deterministic physical–chemical models of the underlying dynamics, will require integration of different data types and models. I suggest two areas poised for progress: quantitative integration of deformation and seismicity; and model-based forecasts conditioned on estimates of material parameters and initial conditions from inversion of available datasets. Deformation and seismicity are the principal geophysical methods for volcano monitoring, and in some cases have signalled dyke propagation minutes to hours prior to eruptions. Quantitative models relating these processes, however, have been lacking. Modern theories of seismicity rate variations under changing stress conditions can be used to integrate deformation and (volcano–tectonic) seismicity into self-consistent inversions for the spatio-temporal evolution of dyke geometry and excess magma pressure. This approach should lead to improved resolution over existing methods and, perhaps, to improved real-time forecasts. The past few decades have also witnessed a marked increase in the sophistication of physical–chemical models of volcanic eruptions. I review conduit models that can be combined with GPS and extrusion rate data through Markov Chain Monte Carlo (MCMC) inversion to estimate the absolute volume of the crustal magma chamber, initial chamber overpressure, initial volatile concentrations and other parameters of interest. The MCMC estimation procedure can be extended to deterministic forecasting by using the distribution of initial conditions and material parameters consistent with available data to initiate predictive forward models. Such physics-based MCMC forecasts would be based on all knowledge of the system, including data up to the current date. The underlying model is completely deterministic; however, because the method samples initial conditions and physical parameters consistent with the given data, it yields probabilistic forecasts including uncertainties in the underlying parameters. Because there are almost certain to be effects not factored into the forward models, there is likely to be a substantial learning curve as models evolve to become more realistic.