本日選文（共 4 篇）。

[907] 利用新建立的 N-net 海底地震儀網於南海溝日向灘檢測淺部構造性顫動

Detection of Shallow Tectonic Tremors in Hyuga‐nada, Nankai Trough, Using the Newly Established N‐Net OBS Network

  • 期刊：JGR: Solid Earth (AGU/Wiley) — eTOC
  • Published：Tue, 17 Mar 2026 02:06:18 -0700
  • DOI：10.1029/2025GL121198
  • 原文連結：連結

Summary (EN) Using machine-learning on newly deployed offshore N‑net OBS data, the authors detected shallow tectonic tremors in Hyuga‑nada between July 2024 and September 2025. Two major tremor episodes occurred: one immediately after the August 2024 M7.0 earthquake and another three days before the January 2025 M6.7 event, with several tremors migrating alongside seismicity—suggesting slow‑slip transients promoted stress transfer along the plate boundary. Tremors clustered in fluid‑rich plate‑boundary zones and were absent in drained areas, indicating structural heterogeneity controls shallow slow‑earthquake dynamics and providing new insights into slow‑earthquake mechanics.

重點摘要（繁中） 作者將機器學習流程應用於新建立的N‑net海底OBS資料，於2024年7月至2025年9月間偵測到日向灘的淺層構造顫動。兩次主要顫動事件分別在2024年8月7.0級地震之後立即發生，以及在2025年1月6.7級地震前三天發生，且多起顫動與地震同時遷移，顯示慢滑移瞬態促進了板塊邊界上的應力轉移。顫動集中在富流體的板界帶並在排水區域不活躍，表明結構異質性調控了日向灘的淺層慢地震動力學，並提供對慢地震機制的新見解。

[915] 斷層異質性增加地震前兆的複雜性：來自含聲發射活動之直接剪切試驗的啟示

Fault heterogeneity increases the complexity of earthquake precursors: Insights from direct shear tests with AE activity

  • 期刊：Earth and Planetary Science Letters (ScienceDirect)
  • DOI：10.1016/j.epsl.2026.119979
  • 原文連結：連結

Summary (EN) The authors report acoustic emission (AE) measurements from laboratory direct-shear tests on discontinuous faults with varying fractions of rock bridges to study how geometrical fault heterogeneity affects seismic precursors. They find that AE event rate and the b-value of AE size distributions delineate progressive failure phases, and identify two precursor anomalies controlled by fault geometry: a power-law acceleration of AE rate before shear localization, and a fluctuating decrease of b-value during shear localization and dynamic failure. The results indicate that the scale of fault heterogeneity strongly influences rupture properties and precursory AE activity, implying that heterogeneity makes earthquake forewarning more complex because of successive damage and off-fault macro-fractures prior to full failure.

重點摘要（繁中） 作者在具有不同岩橋比例的不連續實驗斷層上進行直接剪切試驗並記錄聲發射（AE），以探討幾何不均質對地震前兆的影響。研究顯示，AE 事件率和 AE 規模分佈的 b 值能有效區分逐步破裂階段，並發現兩種由斷層幾何決定的前兆異常：剪切定位前的 AE 事件率呈冪律加速，以及在剪切定位和動態破裂階段出現的 b 值波動性下降。結果表明不均質尺度顯著影響破裂性質與前兆 AE 活動，並指出由於連續損傷與斷層外大尺度破裂，斷層不均質使地震預警更加複雜。

[936] 遠震地震震源機制揭示的海洋轉換斷層幾何、構造與構造體制

Geometry, structure and tectonic regime of oceanic transform faults revealed by teleseismic earthquake focal mechanisms

  • 期刊：GJI (OUP) — RSS
  • Published：Mon, 23 Feb 2026 00:00:00 GMT
  • DOI：10.1093/gji/ggag078
  • 原文連結：連結

Summary (EN) The study uses a global statistical analysis of teleseismic earthquake focal mechanisms from 150 oceanic transform faults and introduces “stack maps” to quantify fault dip and rake and visualize average focal mechanisms. Results show that although many OTFs conform to the classical vertical strike-slip model, nearly half exhibit deviations in dip or motion, which the authors classify into four categories (standard, non-vertical with transtension/transpression, non-vertical strike-slip, and vertical with vertical motion). They find that tectonic regime shifts drive these structural changes and that non-vertical geometries can persist as a structural memory for several tens of millions of years, demonstrating the value of statistically analyzing global seismic catalogs for understanding OTF geometry and tectonic regimes.

重點摘要（繁中） 本研究對150條海洋轉形斷層的遠震震源機制進行全球統計分析，並提出以「堆疊圖」（stack maps）量化斷層傾角與滑動方向，呈現平均震源機制。結果顯示雖有許多OTF符合傳統垂直走滑模式，但近一半斷層在傾角或運動型態上偏離標準，可被歸為四類（標準類、帶有伸張/擠壓分量的非垂直類、非垂直的走滑類、以及具有垂直運動分量的垂直類）。研究指出構造圈域位移的改變是結構變化的主要驅動力，且非垂直幾何形態可作為數千萬年等級的構造記憶，顯示以統計方法分析全球地震目錄能有效揭示OTF的幾何與構造制式。

[972] 從機器學習視角快速且有效地分類滑坡微震性活動

Fast and effective classification of landslide microseismicity: a machine learning perspective

  • 期刊：Earth-Science Reviews (ScienceDirect)
  • DOI：10.1016/j.enggeo.2026.108698
  • 原文連結：連結

Summary (EN) The paper compares supervised and unsupervised machine-learning methods for classifying landslide-related microseismic events from a challenging single-station dataset collected over six months at Bossea Cave (NW Italy). A CNN trained on limited labeled data and using spectrogram inputs plus Monte Carlo Dropout slightly outperforms a PCA-reduced k-means clustering approach, which is strongly affected by seasonal spectral variability from temperature and precipitation and yields more diffuse clusters. Both methods provide satisfactory classification, but the CNN is more robust under non-stationary spectral conditions, highlighting both the promise and limitations of automated single-station monitoring workflows in data-limited operational scenarios.

重點摘要（繁中） 本研究比較監测單站資料下，監督式與非監督式機器學習在地表塌陷微震事件分類的表現，資料來自義大利西北部Bossea洞穴六個月的被動地震觀測。採用頻譜圖並加入蒙地卡羅丟棄法量化不確定性的卷積神經網路，在有限標註資料上略優於以主成分分析降維後的k-means分群；後者受溫度與降雨導致的季節性頻譜變化影響較大，群集呈現重疊與擴散。研究指出兩種方法均能達到可接受的分類效果，但CNN在非定常頻譜條件下更為穩健，突顯單站自動化監測在資料有限情境下的潛力與侷限。