應用地球同步衛星資料於臺灣離島夜間海霧偵測與長期時空變遷分析

Abstract

霧的發生會使得水平能見度降低，並造成交通事故發生的機率增加，因此霧偵測對於提供準確與即時的氣象預報至關重要。使用人工目視與能見度儀為傳統霧偵測方法，然而，這些方法僅能取得單點之資訊。相較之下，衛星遙測可以提供大範圍的資料，但多數的繞極軌道衛星重訪週期為數天，無法進行時間連續的觀測。而地球同步衛星克服了以上缺點，不僅可取得空間連續的資料，且時間解析度高，數十分鐘即產生一筆觀測，在霧偵測上展現出極大潛力。但霧與低雲擁有相似的光譜特徵，使得兩者經常被誤分類。為了更精確地進行霧分類，本研究使用Himawari衛星資料，建立了一種根據光譜與紋理特徵進行分類的機器學習模型，來進行臺灣離島夜間的霧偵測。並且根據2016至2025年的霧偵測結果，分析霧事件長期時空變遷。結果顯示，本研究所建立的海霧模型偵測率（Probability of Detection，POD）達81.66%，誤報率（False Alarm Rate，FAR）為20.76%。相較於只有使用光譜與亮度溫度差異特徵，加入紋理特徵後，模型的POD提高了3.04%。此外，進一步使用遞迴特徵消除法進行特徵選擇，使得模型的POD提升4.82%。在霧時空變遷結果中顯示，夜間海霧頻率較高的範圍集中臺灣海峽北緯24度至25度左右，且澎湖與金門附近海域的海霧頻率高於馬祖區域，並且在10年以來，霧頻率在三個區域未有顯著變化。本研究證明了使用光譜與紋理特能有效提升霧偵測的精度。並且透過長期的霧事件偵測結果，洞察臺灣離島區域霧的時空分布與趨勢，為未來交通安全與規劃工作提供參考。

Fog is a phenomenon that reduces horizontal visibility, often leading to traffic accidents on land and at sea. Therefore, fog detection is crucial for providing accurate and real-time weather forecast guidance. Geostationary satellites have high temporal resolution and spatially continuous observations, showing great potential in fog detection. However, low clouds and fog are often misidentified because of their similar spectral features. To distinguish between fog and low clouds, this study develops a machine learning algorithm based on the spectral and textural characteristics, using Himawari data, for nighttime fog detection in Taiwan’s offshore islands. The research then analyzes spatiotemporal fog patterns and long-term trends based on the results of fog detection from 2016 to 2025. Results show that the probability of detection (POD) and the false alarm rate (FAR) of the model are 81.66% and 20.76%, respectively. Compared to using only spectral features, adding texture features improves POD by 3.04%. Furthermore, applying recursive feature elimination for feature selection increases POD by 4.82%. The spatial distribution of fog indicates that the highest fog occurrence is concentrated in the Taiwan Strait between 24°N and 25°N. Among the offshore islands regions, higher fog frequencies are identified around Penghu and Kinmen than around Matsu. Over the past decade, fog frequency in these regions shows no statistically significant trend. This study improves the accuracy of fog detection and reveals the spatiotemporal patterns of fog in Taiwan’s offshore islands, which supports relevant traffic planning.