使用統計模型探討阿里山茶區夜間成霧原因

Abstract

台灣中海拔山區常受到霧影響並降低能見度，造成交通上的風險，並改變輻射條件與微氣候，進而影響山區農業。本研究以嘉義縣竹崎鄉石園（SY）測站為主要研究站點，並以南投縣溪頭（XT）通量塔為參考站點，探討阿里山周遭茶區的夜間霧頻率為何會明顯高於相近海拔高度的溪頭測站。本研究以每小時通過 QC 之樣本中有霧樣本所占比例定義起霧頻率，兩站霧頻率皆在當地時間約 17 時達到峰值，石園的霧頻率約可達 0.2，溪頭則約為 0.15。相較之下，夜間霧頻率出現很明顯的差異，石園的霧頻率在 20 時由 0.16 逐漸降至凌晨 02 時的 0.08，而溪頭在 20 時後不久即降至 0.05 以下，使得夜間的霧頻率出現三至五倍的差異，兩站霧頻率比值（SY/XT）通常超過 3，並在 22–00 時可達 4.8 以上。 我們建立機器學習模型，用於石園的起霧偵測（t0h）與一小時預報（t1h），以 Logistic Regression 作為基準模型，並以基於決策樹的統計模型（Random Forest 與 LightGBM）作為主要分析重點。SHapley Additive exPlanations特徵重要性診斷（SHAP）提供具有物理可解釋性的線索，用以說明氣溫、混合比、風速、氣壓等氣象條件如何共同影響霧的發生。 為了解釋其環境控制因子，我們根據模型的SHAP分析，挑選較為重要的變數，並檢視其二維霧頻率分布。首先，分析溫度與混合比對霧頻率的關係，結果顯示接近飽和的溫、濕條件可視為成霧的必要條件。其次，在接近飽和的前提下分析平行坡向風速與降溫趨勢對霧頻率的關係，發現上坡風在夜間扮演成霧的關鍵角色。石園夜間（20 – 02時）有霧時的平行坡向風速逐小時中位數約為 0.8 至 1.0 m s⁻¹，無霧時約為 −0.3 至 −0.6 m s⁻¹，其中正值定義為上坡風，負值代表下坡風，故此結果支持夜間上坡風的重要性。溪頭夜間有霧時平行坡向風速的中位數介於 −0.2 至 0.2 m s⁻¹，而無霧時約為 −0.7 至 −0.6 m s⁻¹，與溪頭在 20 時後幾乎缺乏上坡風的現象一致。綜合而言，本研究使用機器學習對霧的存在進行偵測與預報，並指出上坡風對阿里山茶區夜間霧頻率有明顯關聯，季節性的背景風向差異亦可能影響石園夜間風場型態，但其具體作用機制仍待後續研究釐清。

Fog frequently affects the mid-elevation mountainous regions of Taiwan, reducing visibility, increasing traffic hazards, and altering radiation conditions and microclimates, thereby influencing mountain agriculture. This study focuses on the Shiyuan (SY) station in Zhuqi Township, Chiayi County, as the primary study site, with the Xitou (XT) flux tower in Nantou County as a reference site, to investigate why nighttime fog frequency in the tea-growing region surrounding Alishan is markedly higher at SY than at XT, despite their similar elevations. Fog frequency is defined here as the proportion of foggy samples among all quality-controlled samples within each hour. At both sites, fog frequency reaches its peak at approximately 17:00 local time, with values of about 0.20 at SY and 0.15 at XT. In contrast, nighttime fog frequency differs substantially between the two sites. At SY, fog frequency gradually decreases from 0.16 at 20:00 to 0.08 at 02:00, whereas at XT it drops below 0.05 soon after 20:00. As a result, nighttime fog frequency differs by a factor of three to five between the two sites, with the fog frequency ratio (SY/XT) generally exceeding 3 and reaching above 4.8 during 22:00–00:00.

Machine learning models were developed for fog detection at the current hour (t0h) and one-hour-ahead prediction (t1h) at SY. Logistic Regression was used as the baseline model, while decision tree-based statistical models, including Random Forest and LightGBM, served as the main focus of the analysis. SHapley Additive exPlanations (SHAP)-based feature importance diagnostics provide physically interpretable insights into how meteorological conditions, including air temperature, mixing ratio, wind speed, and pressure, jointly influence fog occurrence.

To identify the environmental controls of fog formation, we selected the more important variables based on the SHAP analysis and examined their two-dimensional fog frequency distributions. First, the relationship between temperature, mixing ratio, and fog frequency indicates that near-saturated thermodynamic conditions can be regarded as a necessary condition for fog formation. Second, under near-saturated conditions, the relationship between slope-parallel wind speed, cooling tendency, and fog frequency shows that upslope flow plays a key role in nighttime fog formation. During nighttime (20:00–02:00) foggy conditions at SY, the hourly median slope-parallel wind speed is approximately 0.8 to 1.0 m s⁻¹, whereas under non-foggy conditions it ranges from about −0.3 to −0.6 m s⁻¹. Positive values are defined as upslope flow, and negative values indicate downslope flow, supporting the importance of nighttime upslope flow. At XT, the median slope-parallel wind speed during nighttime foggy conditions ranges from −0.2 to 0.2 m s⁻¹, while under non-foggy conditions it is about −0.7 to −0.6 m s⁻¹, consistent with the near absence of upslope flow after 20:00 at XT. Overall, this study applies machine learning to detect and predict fog occurrence and further indicates that upslope flow is strongly associated with nighttime fog frequency in the Alishan tea-growing region. Seasonal differences in background wind direction may also influence the nighttime wind characteristics at SY, although the underlying mechanism remains to be clarified in future work.