紅邊波段於土地覆蓋分類效益探討

遙測影像應用在土地覆蓋分類是因為每一種土地覆蓋對各個光譜特徵有不相等的反射率。傳統衛星影像由紅光、綠光、藍光與近紅外光波段組成，而近年來有一較新的感測器─紅邊，其波長介於紅光波段與近紅外光波段，該範圍是植物反射率快速變化之區域，因此紅邊波段之設計目的是分出土地上不同種類的植物，達到監測森林之健康程度或是預估植物產量等用途。本研究使用2015/10/16與2015/11/17之RapidEye衛星影像，影像包含藍光波段、綠光波段、紅光波段、紅邊波段與近紅外光波段，進行主成分分析之因素探討、植生指標分類與影像分類，探討紅邊波段在土地覆蓋分類上之效益，針對不同的方法有不同的案例，各個案例成果經由指標評估其差異，進而了解紅邊波段之效益。本研究在主成分分析中先使用全區影像進行主成分分析，接著分別針對水體、裸露地與植物進行主成分分析，以了解每一波段在各種地表覆蓋下之因素負荷值大小，判斷各波段在分類各類地表覆蓋之重要性。全區實驗成果每個波段皆有顯著的貢獻；植物區實驗成果為紅邊波段的因素負荷值僅小於近紅外光，其重要性僅次於近紅外波段。植生指標分類利用六個不同的寬波段與窄波段之植生指標，並將其中三個植生指標(NDVI, CMFI與OSAVI)的近紅外波段換成紅邊波段進行分類，所以總共九個植生指標。植生指標分類有兩種分類成果，第一種是分成植物與非植物，一共兩類；第二種是將植物分成水稻與落花生，將非植物分成水體與裸露地，一共四類。第一種分類成果為，除了CMFI的OA為59%與用紅邊波段替代近紅外波段的OSAVI整體分類精度為84%，其餘植生指標之分類OA高達95%；第二種分類成果OA都在60%~70%，雖然使用紅邊波段代替近紅外波段分類，使NDVI, CMFI和OSAVI的OA降低，檢定測試也顯示有明顯不同，如此，認為紅邊波段無法提升分類精度與分類能力不如近紅外波段，但若從較微觀的角度評估—各類別精度，裸露地精度降低15%、水稻精度提升11%、落花生精度提升7%。而MCARI(紅邊波段植生指標)分類成果能與NDVI與CMFI(非紅邊波段指標)相媲美，從以上成果認為紅邊波段要針對特定的地表覆蓋，才能發揮作用。影像分類部分，使用非監督式(ISODATA與Kmeans)與監督式分類(SVM與ML)對兩實驗區進行分類，第一次使用扣除紅邊波段的RapidEye影像，第二次使用完整的RapidEye影像，分類的類別數有三類(水體、裸露地與植物)與四類(植物分成水稻與落花生或是樹木與作物)。分成三類時，每一次分類的OA都超過80%，分為四類時，作物區OA只有60%~80%，山區OA為70%~90%，紅邊波段作為一特徵進行影像分類能提升1% ~ 3%總體分類精度，Kappa從0.66提升到0.69 (ML)、從0.73提升到0.75 (SVM)經過統計測試，為顯著。本研究主要探討有無紅邊波段影像對於土地覆蓋分類之影響，經由本研究測試發現無論在主成分分析、植生指標分類或是影像分類上，紅邊波段對於地表覆蓋大類分類均有統計上顯著之貢獻。但是，若需對植被做更細一步的區分，則資訊量仍為不足。

關鍵字

主成分分析；植生指標；高斯最大四然概似法；支持向量機

並列摘要

Because each kind of land cover has different reflectance spectrum, remote sensing images could provide rich information for land use/ land cover classification. Traditionally, satellite optical images has red, green, blue and near infread (NIR) bands. Recently, a new band named Red Edge is included. The wavelength of Red Edge is between red and NIR which is the region of rapid change in the vegetation spectrum. The objective of including Red Edge band is to provide more information for applications such as distinguishing different species of vegetation, health onitoring of forest, estimating crop yield. The images utilized in this study are RapidEye satellite images collected on 2015/10/16 and 2015/11/17. The images include blue, green, red, Red Edge and NIR band with the range of wavelength in 410nm-510nm, 520nm-590nm, 630nm-685nm, 690nm-730nm, and 760nm-850nm, respectively. The analysis includes three schemes, Principle Component Analysis (PCA), Vegetation Index (VI) and image classification, for evaluating the effectiveness of Red Edge band on land cover classification. There are different cases for each scheme. The results are quantatively assessed with indices for evaluating the effectiveness of the red-edged band. In order to understand the loading factor and the importance of each band on classifying land cover, PCA scheme has two cases, the whole area, and indivial classes including water body, bare land and vegetation. The result show that every band has significant contribution in the whole area case. However, red band’s loading factor is higher than red edge in plant area, i.e., red band’s contribution is greater than red edge in plant area. There are six indices in VIs classification utilized. Both narrow band indices and board band indices are included. Red edge band replaces red band in the index of NDVI, CMFI, and OSAVI. Therefore, NDVI, CMFI, and OSAVI have two versions in this study. There are nine indices in VIs classification in total. The first case has two classes, vegetation and non-vegetation. The second case has four classes, rice and peanut for vegetated area, and water body and bare land for non-vegetated area. In the first case, OA of most indices is above 95%, except CMFI (OA: 59%) and substituted red edge for red band in OSAVI (OA: 59%). In the second case, OAs are between 60% and 70%. When red edge band replaced red band in NDVI, CMFI and OSAVI, OA is significantly lower from statistical analysis. This indicates that red edge band is unable to promote OA and the classification ability is inferior to red band. However, observing the classification accuracy of each class, bare land decreases 15%, rice increases 11% and peanut increases 7%. The classification result from MCARI (red edge band VI) is comparable with NDVI and CMFI (non- red edge band VI). Summarizing from the above results, red edge band could be helpful for identifying specific land cover types. In the third scheme, image classification methods include unsupervised classification (ISODATA and Kmeans) and supervised classification (ML and SVM). For evaluating the contribution of red edge band, classification with and without red edege band are compared. There are two cases, three categories (waterbody, bare land and vegetation) and four categories (vegetation were divided into rice and peanut or tree and crops). The OA of each classification exceed 80% with three categories. OA in crops area is only 60% to 80% and OA in mountain area is 70% to 90% with four categories. When red edge band is included, OA improved by 1% to 3%, Kappa improved from 0.66 to 0.69 (ML) and 0.73 to 0.75 (SVM). Statistical testing valided that the differences are significant. The purpose of this study is to evaluate the impact of red edge band on land cover classification. The experiments show that red edge band has signification contribution in land cover classification. However, the amount of information is still insufficient for detailed classification in vegetation.