提高周遭噪訊監測的時間解析度與準確性: 應用於2018年台灣東部Mw 6.4的花蓮地震

Abstract

台灣東部的花蓮地區位處於易受頻繁地震威脅的聚合板塊邊界，2018年02月06日規模6.4的花蓮地震是近年造成較多人員傷亡的地震之一，本研究利用連續噪訊記錄結合尾波干涉法量測2017-2018年該區域的地殼速度擾動量，分析其時間與空間的變化分布並探討可能的成因。為了使測量的速度擾動(δv/v)變化 時間解析度能盡量短達數天之內，且同時維持其準確性與穩定性，本研究透過三個程序來達到此需求。首先，計算每天由三分量連續噪訊記錄組合配對的9分量互相關張量(cross-correlation tensor, CCT)；第二，為了能夠以最少的疊加天數重建現時格林函數(current Green’s function, CGF)，測試DOST和SVDWF兩種去噪濾波，作用於單天的CCTs和兩年CCT疊加平均的參考格林函數(reference Green’s function, RGF)，抑制兩者波形的不一致性；第三，使用6種分別在時間域、頻率域與小波域常見的到時偏移測量方法，將尾波波形透過線性拉張、互相關或是非線性最佳擬合(dynamic time warping, DTW)量測CGF和RGF在次微震(secondary microseism)主要頻帶0.1-0.9 Hz區間的走時偏移(δτ)，並評估各方法的表現。 比較未去噪和去噪處理的結果，發現經去噪濾波後，尤其是採用SVDWF，明顯地提高了單天CCT和RGF之間的波形一致性，進而在疊加天數為5天時就快速收斂得到穩定的CGF，獲得可信的 δv/v 估計值。六個方法中，小波交叉頻譜法(wavelet cross spectrum, WCS)測量的 δv/v 擾動振幅在不同站間距和去噪與否的情況下都最為穩定。WCS和DTW兩種方法都具有良好的時間辨析度，因此相較於移動時窗互相關法(moving-window cross correlation)和移動時窗交叉頻譜法(moving-window cross spectrum)，在以線性回歸量測 δv/v 時具有更多的 δτ 控制點，所以不確定性較小。在站間距較小時，波形拉張法(trace stretching)和小波波形拉張法(wavelet trace stretching)也可獲得相對穩定的 δv/v，但當站間距太長就會出現過大的 δv/v。 所有測站的 δv/v 時間序列可以觀察到兩個主要現象，在花蓮地震發生當天突然的同震速度下降，隨後因震後回復而呈現指數回復趨勢。另一特徵是以一年週期主導的準週期性(quasi-periodic)擾動，與降雨所引起的孔隙壓力變化具高度相關性。站間距最長的測站對位於中央山脈東翼且遠離米崙斷層(Milun fault, MF)和嶺頂斷層(Linding fault, LF)，同震速度變化的側向分布除了該測站對以外的其它測站對都顯示同震速度下降。HWA測站位於花蓮市MF南段的上盤，與該測站配對，站間路徑大致呈南北向平行海岸線的三個測站對呈現最為顯著的同震速度下降。這些路徑穿過MF南段和LF北段，附近區域在有限斷層模型滑移量推算的同震靜態體積應變分布呈現伸張應變增加，並且GPS測量得到的地表最大位移(peak ground displacement, PGD)和強地動的地表最大速度(peak ground velocity, PGV)也顯示有較強的振幅。同震速度下降數值以MF為中心呈南北向分布，以此為軸線逐漸往內陸遞減，在朝向內陸的路徑上較不明顯，與PGD和PGV的分布一致。因此我們推斷地殼深度3公里處所觀察到的同震速度下降可能是受到靜態體積應變伸張增加和瞬間動態應力導致地殼力學強度減弱的影響。 不僅是淺層地殼速度擾動，MF沿線上的地下水位和地表位移也具有突然的同震變化和震後逐漸恢復的趨勢，以黏滯力回復的指數函數型態來模擬速度回復的特徵時間，δv/v 和地表位移觀測的最佳擬合顯示回復時間約為1個月，但地下水位的回復時間僅16天，這表示 δv/v 震後變化趨勢主要是受到震後應力回復主導，而地下水位的變化則可能是通過地振動產生的孔隙快速地由含水層補給回復至靜水壓狀態所造成。

We employ noise-based coda wave interferometry to detect disturbance of crustal velocities in time and space and investigate its physical causes in the Hualien area of eastern Taiwan which is located in a plate convergent boundary and highly prone to frequent seismic risk. The investigation period spans from 2017 to 2018 when the deadly Hualien earthquake of magnitude 6.4 struck this area on 6 February 2018. To make the temporal resolution of our estimated velocity perturbations (δv/v) within a few days comparable with geodetic and other geophysical observations, and meanwhile retain their accuracy and stability, we exploit three aspects to meet our need. First, we compute the daily 9-component cross-correlation tensor (CCT) using continuous ambient noise recorded at pairs of the six three-component seismograph stations in the study region. Second, to minimize the number of consecutive days of CCT stack for reconstruction of the robust current Green’s function (CGF), the two denoising filters, named in short as DOST and SVDWF, are tested and exerted to suppress incoherent noise between the daily CCTs and two-year stacked CCT (hereinafter referred to as the reference Green’s function, RGF). Third, we evaluate the performance of the six methods commonly used for measuring the time-lapse shift (δτ) of coda at any given time-lag instant (τ) between the current CGF and RGF in the dominant secondary microseism frequency band of 0.1-0.9 Hz such that the negative slope of the least-squares linear regression of δτ versus τgives the best-estimated δv/v. The measurement of δτ as a function of τ is either made in time, Fourier (frequency), or wavelet domain, replying on linearly stretching, cross-correlating, or non-linear optimally matching (dynamic time warping, DTW) coda traces. Comparing the results obtained with and without denoising, it is evident that the denoise filtering, particularly SVDWF, improves markedly the waveform coherence between each daily CCT and RGF which leads to a rapid convergence to the stable CGF up to five-day stack and thus the reliable δv/v estimate as well. Among the six methods, the δv/v measured by the wavelet cross-spectrum (WCS) method is most consistent and stable, regardless of the inter-distance of the paired stations, the amplitude of the perturbed velocity, and the kind of the denoising filters or non-denoising. Both the WCS and DTW methods have a good localization in time, and therefore more control points of δτ are readily measurable for linear regression of δv/v with a smaller uncertainty, compared to that measured by the moving window cross correlation (MWCC) and cross-spectrum (MWCS) methods. For the pair with a shorter inter-station distance, the trace stretching (TS) and wavelet trace stretching (WTS) methods also produce relatively stable δv/v; however, the excessively large δv/v fluctuation would emerge whenever the inter-distance is too long and/or the velocity perturbation is too subtle. The temporal δv/v variations observed at all the stations reveal the two main features. One exhibits a sudden coseismic velocity drop on the date when the Hualien earthquake occurred, succeeding an exponential recovery due to postseismic viscous relaxation, while the other displays quasi-periodic fluctuations with a dominant annual period correlated well with the precipitation-induced pore pressure change. The lateral distribution of coseismic velocity change shows that the velocity decrease is omnipresent for all the station pairs except for the one of the longest inter-station distance mostly through the eastern limb of the Central Range away from the Milun fault (MF) and Linding fault (LF) zones. The most pronounced velocity drop occurs along the nearly N-S trending, coastline-parallel paths paired with station HWA, which is located in Hualien City on the hanging wall of the southern segment of the MF. These paths all traverse the southern MF and northern LF segments, where the nearby region experiences the greatly enhanced coseismic static dilatational strain induced by the slip distribution of the finite-fault model and the much stronger amplitudes of the GPS-derived peak ground displacement (PGD) and the peak ground velocity (PGV) inferred from seismic shaking. The coseismic velocity reduction becomes gradually less visible for the paths toward inland, consistent with the N-S elongated concentric distribution of the PGD and PGV values centered around the MF and outwardly decreasing in strength. We therefore deduce that the increased static dilatational strain and the mechanical properties weakened by transient dynamic stresses in the uppermost 3 km of the crust are responsible for our observed coseismic velocity reduction. Not only δv/v but also the surface displacement along the MF and ground water level possesses the sudden coseismic change and gradual postseismic recovery, which we model as the sum of a constant term for the coseismic velocity reduction and an exponential term with a characteristic time of viscous relaxation for the transient velocity recovery. The relaxation time taken from the best-fit models is about 1 month for δv/v and the surface displacement, but shortened to 16 days for the ground water level. This finding suggests the δv/v recovery is primarily governed by postseismic stress relaxation through viscous creep. However, the faster recovery of the excessively reduced pore pressure may imply the hydrostatic condition at the water table as a result from the rapid aquifer recharge through the shaking-induced fractured conduit.