物件導向分析方法應用於遙測影像之分區及崩塌地與人工設施分類

Abstract

對於大範圍的受災情況，利用遙測影像能在短期內獲得大區域災情資料，但藉由傳統人工判釋與數化方法對於廣域影像判釋分類頗為耗時，可能影響即時與快速決策。如何利用自動判釋方法快速進行影像上之地物分類是相當重要的課題:過去常用之自動判釋方法大多採用像元式的分析方法，然而像元式分類法僅藉由光譜分布的差異進行分類，不易呈現地物間之空間關聯性，且分類結果易有雜訊產生，影響分類的精度和結果。 本研究利用物件導向式遙測分析方法，使用福爾摩沙二號衛星在莫拉克颱風事件前後遙測影像做大範圍坡地與河川災害快速判釋。本研究使用災害前後期影像同時進行均質化分割技術，於兩張影像上產生出相同區塊，以解決不同時期影像之地貌分類邊界問題；以區塊為分析單元，依照人工判釋經驗與階層式的邏輯，加入合適之空間特徵資訊（包括形狀、光譜值、坡度、空間關係），訂定規則流程，由單純至複雜的地貌進行十五種分類，建立出系統化之分類程序。 應用於研究區之分類結果，進一步與航照判釋及現地踏勘比對，最後訓練成果由誤差矩陣評估得整體精度達85.2%。研究案例地貌分類結果於研究區域顯示：崩壞比由災前的1.2% 增加為8.9%、河道變遷河段達690處，災後河道面積相對於災前增加40%、受影響之開發地達75%、受影響之房舍達37%（1268處）、受影響之道路達30%。

In this study, object-oriented analysis method was applied to interpret landslide and flood disasters before and after Typhoon Morakot with remote sensing images. It is shown that remote sensing images in wide area can be recognized quickly using this method. Two images before and after the disaster were homogeneously segmented to generate the same blocks in order to solve the boundary problem in feature classification. Based on the rules of artificial interpretation, proper information (ex. shape, spectral values, slopes , Spatial relations…etc.) was added to construct hierarchical logic, classifying 15 different kinds of features. The classification results are double checked with aerial photographs and field investigation. The overall accuracy of final training outcome based on the error matrix assessment is about 85.2%. The results show that the landslide ratio increased from 1.2% to 8.9%, blocks features of river channel increased 690 sites, area ratio increasing to 40% compared with the original river channel area, 75% of the developed lands were affected, 37% （1268 sites）of the houses were affected, 30% of the roads were affected .