地震事件偵測及P波與S波抵達時間判定是地震監測與後續震源分析的重要基礎。近年來，基於深度學習的方法雖能取得良好預測表現，但通常需要較長的模型建立時間及較高的計算資源。為兼顧震相判定能力與計算效率，本研究提出具低計算成本與可解釋性的地震偵測與震相判定綠色學習方法。首先對三分量地震波形進行濾波、正規化及重疊視窗切分，再利用兩層PixelHop提取不同層次的局部波形特徵，其中第一層取得直流與交流成分，並將交流成分輸入第二層以獲取更高層次特徵。接著，透過Relevant Feature Test篩選具代表性的特徵，降低特徵維度與冗餘資訊，最後將篩選後的特徵輸入XGBoost迴歸模型，產生各時間點的預測值序列，並據此判定地震事件及P波與S波抵達時間。本研究使用STEAD、CWA Benchmark及台灣地區連續地震波形資料集進行實驗，並與多種現有深度學習方法比較。實驗結果顯示，本研究方法具備有效的地震事件偵測與震相抵達時間判定能力，其中P波的整體判定表現較為穩定，S波亦具有良好的召回能力。在計算成本方面，相較於既有的深度學習方法，本研究方法所需的執行時間、能源消耗量與碳排放量皆明顯較低，即能以較低計算成本完成地震波形分析，在計算效率、能源使用及可解釋性方面展現優勢，可作為深度學習方法之外的另一種選擇。;Earthquake detection and P- and S-wave phase picking are fundamental tasks in seismic monitoring and subsequent source analysis. Although deep learning-based methods have achieved strong predictive performance in recent years， they generally require considerable model development time and computational resources. To balance phase-picking performance with computational efficiency， this study proposes an interpretable and computationally efficient Green Learning method for earthquake detection and phase picking. Three-component seismic waveforms are first processed through filtering， normalization， and overlapping-window segmentation. A two-stage PixelHop framework is then employed to extract local waveform features at different levels. The first stage produces direct current (DC) and alternating current (AC) components， and the AC components are subsequently passed to the second stage to extract higher-level features. The Relevant Feature Test is then applied to select representative features and reduce feature dimensionality and redundancy. Finally， the selected features are fed into an XGBoost regression model to generate a sequence of prediction scores across time points， from which earthquake events and P- and S-wave arrival times are determined. Experiments are conducted on STEAD， CWA Benchmark， and a continuous seismic waveform dataset from Taiwan， and the proposed method is compared with several existing deep learning methods. The experimental results demonstrate that the proposed method can effectively detect earthquake events and determine phase arrival times. Its P-wave picking performance is relatively stable， while its S-wave picking performance also achieves high recall. In terms of computational cost， the proposed method requires substantially shorter execution time and results in lower energy consumption and carbon emissions than existing deep learning methods. Therefore， it enables seismic waveform analysis at a lower computational cost and offers advantages in computational efficiency， energy efficiency， and interpretability， providing an alternative to deep learning-based approaches.
