遙測影像之雲霧偵測及干擾去除

Abstract

遙測技術發展的最大限制之一，便是大氣活動的干擾，尤其對地球資源衛星上裝載可見光段的光學感測器而言，雲霧的干擾最不易處理。而台灣又為多雲區，雲霧的干擾無可避免，若受遮蔽的區域恰為研究區，則會大大減低了衛星影像的價值性。另外，國內災害監測對遙測影像之需求週期幾乎以「天」為單位，更突顯災害發生時遙測影像供不應求之窘境。因此如何去除雲霧之干擾便成為相當重要的課題。本研究即針對上述課題，發展雲霧的偵測及影像鑲嵌之技術。 厚、薄雲霧兩者具有不同之光譜特性，故需個別進行處理。厚雲霧因具高反射特性，故可經由閾值的設定將可見光段呈高反射之部分偵測並去除；薄雲霧雖仍可觀測到底下地物，但已將地物本身的光譜特性扭曲，且難以用演算法偵測去除，因此本研究先將影像由RGB轉換成HIS系統，再假設薄雲霧的加入等於使影像加入白色，即R、G、B三者提高，因此僅改變光譜的亮度或飽和度值，色相並無改變，藉此可偵測並去除薄雲霧。去除厚薄雲霧後殘缺之部分再利用影像鑲嵌之方式，以鄰近日期之無雲影像補償之，並進行色差之調整，使影像資訊量損失到最小。 對於雲霧偵測之結果，本研究以專家法之方式進行檢核。結果厚雲霧部分之總體精度可達97％；在均質海面及有地形效應之區域上，薄雲霧偵測精度約為83％；極度非均質之農業區則降為80％。地物較複雜之區域，需取得較高時間精度之影像才可增加偵測精度。本研究雖因影像取得之限制而無法提升複雜區域之精度，但已證明在HIS系統中可簡化薄雲霧之偵測準則，大大提升自動化偵測雲霧之可能性。

Detection and removal of cloud and haze are arduous problems in optical remote sensing imagery processing. Thick cloud and haze have the character of high reflection, so we can set the threshold to detect and remove the areas having extremely high reflection and even mosaic the images with near dates’ ones to create clear and cloudless images. Relatively, areas covered by thin cloud and haze have the spectral characteristics of both surface features and cloud and haze, thus making it difficult to separate them. This research first processed the images with relative radiometric normalization and then transformed the images from the RGB to the HIS color model. Our assumption was that the interference of thin cloud and haze, similar to mixing a color pigment with white, would increase the color intensity and decrease the saturation of an image but would not change its hue value. Guided by this assumption, we processed the multi-temporal images and isolated areas contaminated by thin cloud and haze. The results thus suggest that an automatic method based on the HIS color model is possible for detecting thin cloud and haze on satellite images.