| 摘要: | 在後疫情時代，許多跨國企業為分散供應鏈風險而轉移工廠回台灣，在工業用地需求增加下，導致部分工廠選擇建設於非工業用地，若缺乏適當地監督及管理，可能威脅鄰近農田或住宅環境。為避免此一情況，定期地監測土地利用及土地覆蓋變化有助於偵測違規土地開發行為。
本研究區域為臺灣北部的桃園市，近年來因都市發展迅速，若以人力進行土地變異查緝曠日廢時，而使用光學衛星影像進行變遷偵測，可借助其高即時性及大面積觀測優勢，提高監測效率，若再透過整合不同來源的衛星影像，將能夠進一步地提高監測頻率。
過去文獻中曾提出許多不同變遷偵測方法，然而對於季節性變化顯著的區域如水稻田等農業區，植物生長周期之地物覆蓋因季節不同而常被誤判為變異區域，成為判斷是否確為變遷的問題。對此，本研究蒐集2017至2021年間涵蓋研究區的Sentinel-2及SPOT-6/7影像，得到長時間跨度的影像序列，將時序資料進行相對正規化處理，並透過最小二乘法擬合週期性變化分析其藍、綠和近紅外光波段反射率在五年內的震盪，從而建立出時間與反射率強度之關係，並基於該成果為不同季節設定門檻值，再透過門檻值與新拍攝之影像進行比較，偵測出新拍攝影像中的異常區域，可降低誤判具有季節性物候區域為變異點的情況。
經與驗證資料分析比對，研究成果能偵測出總數71%之變異點，若以面積計算其整體精度高達99%、Kappa Coefficient達0.86，顯示本研究方法能成功排除誤判季節性變遷的區域，並有效偵測出真實地表變異。
;In the post-pandemic era, many corporations have relocated factories to Taiwan to diversify their supply chain risks. Due to increased demand for industrial land, some factories may build in non-designated zones without appropriate management, threatening land usage in the neighboring residential and agricultural areas. In this situation, periodical monitoring of land use and land cover changes can assist in identifying unauthorized land development. Therefore, our study focuses on monitoring land change in Taoyuan City in northern Taiwan, where urban growth has been accelerated in recent years. Investigating land use change by on-site survey may take a long time and could be more efficient; hence, using optical satellite imagery for change detection has the advantages of high timeliness and extensive observational coverage, improving monitoring efficiency. By integrating satellite images from different sources, the frequency of monitoring can be further increased.
Various methods for change detection have been proposed. Still, for areas with significant seasonal change, such as rice fields and other agricultural zones, changes in the land surface due to plant growth cycles can be misinterpreted as anomalies, posing challenges in assessing surface changes. In the past, various methods have been proposed for change detection. However, regions with the abovementioned ambiguities may be misinterpreted as changes, posing challenges in change detection. In this study, we collected a series of Sentinel-2 and SPOT-6/7 satellite images covering our study area from 2017 to 2021. We process these images through relative radiometric normalization and use a periodical model to fit changes in reflectance (blue, green, and near-infrared bands) in a least squares sense over the five-year time series. This allows us to establish a relationship between time and reflectance, setting a dynamic threshold value for any given time in a year. This model can use these thresholds with a newly captured image to detect areas that had changed in the last epoch, reducing misclassification areas with seasonal phenology as change points. Comparing our results with reference data, our method successfully detects 71% of total change points, achieving an overall accuracy of 99% and a Kappa coefficient of 0.86 in area validation. It is concluded that our method can avoid misclassification from seasonal change areas and effectively detect actual change points. |
