摘 要 台灣全島面積一半以上為國有林地，若發生森林火災，往往造成生態環境之損害，因此快速的森林 火災製圖，便成為一項十分重要之工作。本研究以2009 年大埔事業區所發生的森林火災為研究對象，以 Z/I DMC（Digital Mapping Camera）數值航攝影像為材料，探討火災區域自動化製圖之可行性。具體方式 係採用NDVI 影像差值法(NDVI differencing)、光譜變化向量分析法(spectral change vector analysis, SCVA)、 主成份分析法(principal components analysis, PCA)三種變遷偵測技術，分別結合反覆自我組織分析技術 (iterative self-organizing data analysis technique, ISODATA)非監督分類方法進行受災區域分類，除針對高解 析航攝影像特性逐步去除雜訊外，並探討各偵測技術結合非監督分類之差異性，藉以建立高解析航攝影 像於火災區位自動製圖之適當處理流程。研究結果顯示，各種影像增揚處理方式皆能有效突顯火災區域 之特徵，然而搭配ISODATA 非監督分類法，區分火災受災與非受災區域時，會受到增揚影像的數值分布 所影響；在研究中，增揚影像的數值分布以單波峰偏斜分布，似乎較適合ISODATA 區分兩類。經陰影濾 除、形態學處理後皆可提升分類總體精確度，降低誤差。整體而言，以NDVI 影像差值法搭配陰影去除、 形態學後處理之精度最佳，總體精度達91.93%，且自動化程度高，本研究的森林火災自動製圖技術可供 相關單位應用。

ABSTRACT In Taiwan, more than half of area is national forest land, if there was forest fire which often causes damage to the ecological environment. Therefore, rapid mapping of forest fires have become a very important job. We studied the forest fires of Dapu working cycle which was occurred in 2009, and we used Z/I DMC airborne multispectral image as material. In order to investigate the feasibility of automated forest fire mapping. We use three kinds of change detection techniques that include NDVI differencing (dNDVI), spectral change vector analysis (SCVA), principal components analysis (PCA), and we combined the three of change detection techniques and iterative self-organizing data analysis technique (ISODATA) to classify the fire affected area. We focused on the characteristics of high-resolution image to remove the noise of change detection, and to explore the differences between the three change detection techniques that combination ISODATA unsupervised classification. In order to establish the appropriate processes of automatic forest fires mapping for airborne multispectral image. The results showed that each change detection technique can enhance the characteristics of the fire area effective. However, those change detection techniques with ISODATA to distinguish between burned and unburned area, which will be affected by the distribution of enhanced image values. The numerical distribution of images to a single peak skewed distribution seems to be more suitable for two categories of classification using ISODATA. Image processing of shadow removal or morphology can reduce classification error and improve the overall classification accuracy. Overall, the dNDVI method with image processing of shadow removal and morphology showed the best overall accuracy of 91.93%, and with highly automated processes. The forest fire automatic mapping technology can provide the coherent units to be applied.