| 摘要: | 人們對於空氣品質的意識日漸抬頭，在台灣地區常見的氣膠污染來源可以分成長程傳輸的境外間接污染以及與我們生活中人為活動容易接觸到的一次性污染。為了有效且即時監測小範圍地區污染源流動的幅度與該地區整體時空遭受污染的變異程度，政府、企業甚至是民間單位攜手合作建立起全台的空氣品質監測網。固然地面測站的測量可以更精準的了解一地的影響程度，但儀器本身的保養、校正、管理都需要耗費大量時間與人力，並且考慮到地形上的限制也無法有效提供完整的空間資訊。為此，透過衛星遙測技術結合氣膠反演方法，不僅可以改善監測方法於時間及空間上的缺陷在校正上也可以因為使用同一儀器，無須如地面測站一般各自獨立解決。目前服役的光學衛星中，10公尺空間解析度與每5天為單位的Sentinel-2最為精緻，該衛星是配合歐洲太空總署(European Space Agency, ESA)的計畫來進行，因此可以確保該資料在後續會持續穩定的提供。然而，在Sentinel-2產品應用上卻發現結果大有問題，將其與有相似波段的Landsat-8來做比較，根據輻射傳輸原理發現原因在於Sentinel-2的地表反射率(Land Surface Reflectance)上。因此本研究為了改善Sentinel-2地表反射率之問題，首先會利用地物分類法將地表進行NDVI區分，接著依照得到的區間套用各自的回歸式來做修正，最後則是將原始Sentinel-2大氣層頂反射率(Top-Of-Atmosphere reflectance)資料與修正後地表反射率一同帶入離散係數法中反演氣膠光學厚度，來有效監測地區性的氣膠污染排放程度。
首先利用監督式分類法區分出密集植被區(DVA)、疏散植被區(BVA)、亮區(BA)三種地物下的歸一化植被指數(Normalized Difference Vegetation Index, NDVI)區間，發現NDVI大於0.3的DVA、BVA地區Sentinel-2反射率高估的情形特別明顯，因此帶入自身短波紅外光(SWIR2, 2.19μm)與可見光之間的線性模型修來做修正，至於BA地區則透過Landsat-8大氣層頂反射率與地表反射率的回歸式來處理。在30公尺解析度下的比較，修正後Sentinel-2 AOD落在期望誤差(Expected Error)內的比例藍光達到47%與Landsat-8 AOD的24%相比來的更好；而在10公尺解析度下的比較，雖然與AERONET AOD的相關性還有待修正，但整體絕對誤差與均方根誤差(RMSE)的表現有明顯的改善，說明本研究所提出的地表反射率計算方法可以有相對不錯的可行性。
;Increasing awareness of air quality in Taiwan has highlighted aerosol pollution sources, including long-range transport and local anthropogenic activities. To monitor pollution effectively in real-time, the government, enterprises, and civil sectors have collaborated to establish an air quality monitoring network. While ground-based stations offer precise measurements, their maintenance and geographic constraints limit spatial coverage. Satellite remote sensing combined with aerosol retrieval methods can address these limitations, providing consistent calibration across instruments. Sentinel-2, with its 10m spatial resolution and 5-day revisit time, is particularly suitable for this purpose. However, discrepancies have been found in Sentinel-2 Land Surface Reflectance (LSR) compared to Landsat-8, necessitating improvements.
This study aims to enhance Sentinel-2 LSR by classifying land cover using NDVI and applying specific regression corrections. The corrected LSR and original Top-Of-Atmosphere (TOA) reflectance are used for aerosol optical depth (AOD) retrieval via the Discrete Coefficient Method(DCM). Three land cover types were identified: Dense Vegetation Area(DVA), Barely Vegetation Area(BVA), and Bright Areas (BA). Overestimation in areas with NDVI >0.3 was corrected using SWIR2 and visible band regressions, while BA areas were corrected using Landsat-8 TOA-to-LSR regression.
At a 30m resolution, the corrected Sentinel-2 AOD exhibits a higher percentage within the Expected Error (EE) range compared to Landsat-8. Specifically, 47% of the corrected Sentinel-2 AOD falls within the EE for the blue band, while only 24% does for Landsat-8. At a 10m resolution, despite the need for improved correlation with AERONET AOD, both the absolute error and RMSE showed significant improvement. This validates the proposed LSR correction method for effective regional aerosol pollution monitoring, suggesting its feasibility and potential for broader application. |
