應用遙測與知識庫系統進行林型分類之研究

Using Remote Sensing and Knowledge-Based System to Study Forest Type Classification

陳建文(Cheng-Wen Chen)

指導教授 ： 陳朝圳

國立屏東科技大學/農學院/森林系所/碩士(2010年)

https://doi.org/10.6346/NPUST.2010.00161

摘要

近幾十年來，由遙測影像之空間解析力、時間解析力及波譜解析力的提升，利用衛星影像進行土地覆蓋與土地利用分類，有輔助航空照片之趨勢，且已大量使用於資源調查及環境監測上。台灣地區由於地形起伏變化大，森林結構與樹種組成複雜，因此在遙測影像分類時，如以光譜資訊進行林型分類，仍有其侷限性。本研究目的在於探討如何整合林型的四季光譜資訊與生育環境因子，並尋求最佳的分類方法，以提升分類準確度。本研究利用SPOT衛星影像為材料，以花蓮高山地區為研究範圍。不同季節之衛星影像以直方圖匹配校正法、ATCOR2配合藍氏反射校正法與ATCOR3校正法等三種輻射糾正法，比較大氣與地形效應之糾正效果。另以高海拔地區之主要林型包括冷杉、雲杉、鐵杉、柳杉、檜木、箭竹、松類及裸露地為分類類型，以不同季節之SPOT衛星影像，萃取各林型之均質影像，藉由雙光譜與單因子變異數分析，進行各林型之各波譜之統計分析，利用各林型光譜差異最大之季節，進行最大概似法之監督性分類，以供對照用。為提升分類準確度，本研究以分類迴歸樹演算法(Classification And Regression Tree, CART)結合四季光譜資訊、植生指標及林型之生育環境資訊，利用知識庫分類法演算類樹，進行林型分類，探討知識庫分類法在衛星影像之林型分類的可行性。研究結果
顯示台灣高山地區之SPOT衛星影像，以ATCOR3校正模式能有效的進行地形效應糾正，由不同季節之各林型雙光譜與單因子變異數分析得知，夏季影像對於各林型間之波譜差異性較大，且因太陽直射地面，產生的陰影面積最少，適宜進行林型分類，以夏季影像利用最大概似法進行冷杉等八種林型之監督性分類，其整體分類準確度為62.00%。應用CART所建立之分類樹，將四季影像波譜資訊加入植生指標，進行知識庫分類，其分類準確度提高為66.91%，而將四季影像波譜資訊與植生指標配合林型分布之海拔高之生態幅度資訊，則分類準確度可達到70.18％，顯現知識庫分類如能找到關鍵的波譜以外的分類輔助資訊，確實可提高分類準確度。

關鍵字

分類迴歸樹 ； 知識庫系統 ； 輻射糾正 ； 光譜特徵 ； 植生指標

並列摘要

In last decades, remote sensing images are improved by their spectral, spatial, radiometric abilities, and temporal resolutions. Using satellite images to classify the land cover and land use is the trend and has been widely used in resource surveys and environmental monitoring. Due to the great undulating terrain, the forest structure and species composition are complicated in Taiwan therefore spectral information for forest type classification still has limitations. The purpose of this study is to combine the season’s spectral information and the environmental factors of forest types in order to improve classification accuracy by finding the best classification method. We used SPOT satellite images as a material in the alpine of Hualien. The terrain correcting method of histogram matching, ATCOR2 with Lambertian Reflection and ATCOR3 were used to correct the satellite images in different seasons and to compare the correcting effects of atmosphere and topography. We used the SPOT satellite images of different seasons to classify the major forest types that include Abies kawakamii, Picea morrisonicola,
Tsuga chinensis, Cryptomeria japonica, Chamaecyparis formosen Arundinaria niitakayamensis, pine and bare land in the alpine of Hualien. The bi-spectral plot and one-way ANOVA were used to analyze the greatest difference of the season spectral values by extracting the homogeneous spectral image of different forest types and used maximum likelihood classification to classify forest type for reference as a sample. Next, the knowledge-based systems that integrated the season's spectral information, vegetation indices and forest types of the environmental information were built by Classification And Regression Tree (CART). This study used knowledge-based system to assess the feasibility for forest type classification. The results show that the ATCOR3 correction method is effective for topographic effects of SPOT satellite images in alpine areas of Hualien. The bi-spectral plot and one-way ANOVA analysis indicates that the summer’s image has the greatest difference spectral values from others seasons and the direct sunlight in summer can produce less shadow area. Then, using summer’s satellite images, 8 forest types were classified by the supervised classification of maximum likelihood method, the overall accuracy is 62.00%. We used CART method to integrate the season's spectral information and vegetation indices and to build the knowledge-based system and the results show that the classification of overall accuracy increased to 66.91%. Furthermore we used CART method to integrate the season's spectral information, vegetation indices and forest types of the environmental information to build the knowledge-based system and the overall accuracy of classification was upgraded to 70.18%. In this study, we find the knowledge-based system that can extract the key point of auxiliary information and to improve the classification accuracy.

並列關鍵字

Classification and regression tree ； Knowledge-based system ； Radiometric correction ； Spectral feature ； Vegetation index

參考文獻

陳建璋、吳守從、陳建文、陳朝圳 (2009) 利用SPOT衛星影像探討阿里山事業區林型地物光譜特徵。作物、環境與生物資訊 6: 122-131。

羅時凡、魏浚紘、李崇誠、陳朝圳 (2009) 遙測技術應用於溪頭地區柳杉人工林之森林健康調查。林業研究季刊 31(1): 37-54。

周天穎、簡甫任、雷祖強 (2003) 都市地區土地利用變遷量化分析之研究。台灣土地研究 6(1): 105-130。

呂福原、歐辰雄 (1997) 樹木解說。行政院農委會出版。205頁。

馮郁筑、陳朝圳 (2008) 墾丁國家公園銀合歡植群之光譜反射特徵研究。特有生物研究 10(1): 73-83。