在氣候與環境研究中，如何準確描述空間異質性對於理解複雜的降雨過程至關重要，尤其當降雨資料同時具有大量零降雨值與厚尾分布特性時，更增加了分析上的困難。日降雨資料通常包含大量無雨日，以及少數極端強降雨事件，因此在統計建模與風險評估上具有相當大的挑戰性。Abbas et al. (2025) 所提出的零膨脹延伸廣義帕累托（zero-inflated extended generalized Pareto， ZIEGP）模型，提供一套統一的建模架構，可同時描述零降雨與所有的降雨強度，且不需事先設定極端值門檻。然而，現有的 ZIEGP 模型假設參數在空間上為固定常數，難以充分反映地形、季風系統及區域氣象機制所造成的局部氣候差異。為了克服此限制，本研究建構一個具有空間變異參數的 ZIEGP 模型，並利用多解析度的有序基底函數來表示模型參數的空間變化。此方法除了能彈性地刻畫不同區域的降雨特徵外，也能兼顧大型空間資料分析下的計算效率。由於在複雜空間相依結構下進行完整概似推論具有高度計算負擔，因此本研究採用複合概似估計架構，以提升參數估計的穩定性與計算可行性。此外，為了選取各空間變異參數所需的適當空間解析度，本研究進一步提出複合概似廣義貝氏訊息準則（composite likelihood generalized Bayesian information criterion， CL-GBIC），作為基底函數數量選擇的方法。模擬研究結果顯示，所提出的方法在不同資料生成情境下，皆能有效還原空間變異參數，並選取適當的基底函數配置。最後，本研究將所提方法應用於台灣高解析度日降雨資料分析，結果顯示該模型在空間降雨分析與氣候風險評估上具有良好的實務應用價值。;In climate and environmental studies， accurately characterizing spatial heterogeneity is essential for understanding complex precipitation processes， particularly when rainfall data exhibit both excessive zeros and heavy-tailed behavior. Daily precipitation records commonly contain a large proportion of dry days together with occasional extreme rainfall events， creating substantial challenges for statistical modeling and risk assessment. The zero-inflated extended generalized Pareto (ZIEGP) model proposed by Abbas et al. (2025) provides a unified framework capable of jointly modeling zero precipitation and the entire range of positive rainfall intensities without requiring a pre-specified threshold. Nevertheless， existing ZIEGP formulations generally assume spatially homogeneous parameters， which may be inadequate for capturing localized climatic variability induced by terrain， monsoon systems， and regional meteorological mechanisms. To address this issue， this study develops a spatially varying ZIEGP model in which model parameters are represented through ordered multi-resolution basis functions. This representation allows the proposed model to flexibly characterize spatially varying precipitation structures while maintaining computational efficiency for large spatial datasets. Since likelihood-based inference becomes computationally demanding under complex spatial dependence， a composite likelihood estimation framework is adopted to facilitate stable and scalable parameter estimation. In addition， to determine an appropriate spatial resolution for each varying parameter， this study proposes a composite likelihood generalized Bayesian information criterion (CL-GBIC) for basis function selection within the composite likelihood framework. Simulation studies demonstrate that the proposed methodology can effectively recover spatially varying parameters and identify suitable basis function configurations under various data-generating scenarios. Finally， the proposed approach is applied to high-resolution daily rainfall data in Taiwan， highlighting its practical usefulness for spatial precipitation analysis and climate risk assessment.
