變遷偵測技術在自然資源管理和監控土地覆蓋/使用扮演一個不可或缺的角色，過去十幾年，透過時序性多光譜衛星影像的各種變遷偵測方法已被提出，雖然多數方法足以區分主要的變異，但判釋後的變異點易屬季節變化所致。此外，有些方法需根據反覆試驗或經驗法則來監督特定土地覆蓋/使用類型的門檻值，本研究目標為降低監測區域季節性或長週期反覆變化所造成之誤判，並在無監督式的情況下自動進行變遷偵測。 因此，我們提出了基於一系列工作流程的完全無監督式變遷偵測方法，為了容許植被/作物區的季節性物候在監測區所引起之正常光譜變化，建立2017年至2019年的SPOT-6/7和Sentinel-2的影像資料庫，以識別每個像元的時間特徵，可用來設置新影像的容許值，為了重疊來自不同衛星的每個像元，重新投影和重新採樣為必需的過程，Sentinel-2產品的1C級也需要進行額外的大氣校正，以減少來自不同大氣條件的變化。本研究採用基於虛擬不變點的相對影像正規化技術來調整像元值的範圍，經過以上步驟，提出的方法將每個像元的標準差和平均值記錄在歷史影像資料庫中，如果新影像像元值落在該像元的容許值中，我們視為未變異的像元。 實驗結果透過臺灣宜蘭縣的假色衛星影像（近紅外光-紅光-綠光）來實行，提出的實用方法能檢測突發的變異區，亦可偵測出季節性變化，自我驗證部分，總體準確性為97％，kappa為93％。與現地資料相比，本研究可以減少39％因季節性變化所造成的誤判。 ;The change detection (CD) technique plays an essential role in natural resource management and land cover/use monitoring (LCUM). Over the past decades, various approaches have been proposed for CD by multitemporal and multispectral satellite imageries. Although most of these approaches are enough to distinguish primary change polygons, the recognized changes are prone to seasonal variability. Moreover, some approaches supervise model thresholds by tuning the optimized parameters for a specific land cover/use type. Our purpose is to suppress seasonal variability of detected areas and to implement change detection automatically. Therefore, we propose a fully unsupervised CD method based on a retrospective analysis workflow. In order to suppress extra noise of detected areas caused by the seasonal phenology of vegetated/crop areas, a database consists of SPOT-6/7 and Sentinel-2 imagery from 2017 to 2019 is built to recognize each pixel′s temporal signature, which can be used to set up a tolerant threshold for the coming images. To co-register each pixel from different satellites, reprojection and resample are necessary procedures. Level-1C product of Sentinel-2 also requires additional atmospheric correction to reduce specific changes from different atmospheric conditions. We further adopt the relative radiometric normalization (RRN) technique based on pseudo-invariant features (PIF) to rearrange pixel values. After the above steps, the proposed method records the standard deviation and mean values of each pixel in the historical image database. If the pixel value of the coming image is within the range of tolerance, we recognize it as an unchanged pixel. Experimental results are carried out using false color compositions (NIR-R-G) satellite imageries in Yilan County, Taiwan. The proposed method would be more practical and can detect abrupt change areas. In our preliminary results, the overall accuracy of CD as compared against visual inspection is 97%, with a kappa value of 93%. This workflow can reduce 39% of seasonal changes that are likely to be misidentified from a single pair of images.
