森林蓄積量是森林經營管理決策之重要參考資訊,其資訊的取得係利用現地永久樣區調查,經由樣區資料之處理分析後,提供決策者之參考依據。目前國內外森林資源調查方法,主要以地面樣區調查資料,配合空間性的航測影像,利用雙重取樣的方式,建立空中樣區與地面樣區之推估模式。本研究以屏東林區為例,利用地面樣區調查資料配合相對應位置之遙測指數,建立各不同林型之森林蓄積量推估模式,並探討不同取樣方式對推估模式精度的影響。另利用K最近鄰法,以最短距離選取K個最近鄰樣區,進行森林蓄積量推估。研究結果顯示地面樣區資料可繪製森林資源調查樣區位置圖,並且透過方位角進行樣區位置校正。而利用地面樣區所調查之資料,計算森林蓄積量約為36,531,244 m3,透過分層取樣的方式,可提升森林蓄積量之推估能力,以分層取樣推估之蓄積量約33,956,727 m3。遙測影像配合地面樣區資料,可建立具空間分布之林分蓄積量推估模式,遙測模式推估結果顯示,遙測植生指數於本研究中,RVI、DVI、GRVI具備較佳之森林蓄積量推估能力,其 RMSE = 61.71 m3/ha。K最近鄰法於蓄積量推估時,可以結合地理空間變數(海拔高、坡度、坡向),可獲得較低的K值與RMSE。而利用單一遙測指數變數(NDVI)進行推估,其最佳K值為 14。但透過上述所有4種變數的結合,能將K值降低至6,RMSE可降低至101.02 m3/ha,證明透過多項變數導入,能改善K最近鄰法對於森林蓄積量推估之效果。以本研究而言,遙測模式推估森林蓄積量之成果較K最近鄰法佳,推測原因為本研究之地面樣區數量,以K最近鄰法無法取得最佳之蓄積量推估結果,而當地面樣區數量足夠時,以K最近鄰法推估空間性之林分蓄積量為ㄧ可行方法。
Forest stocking is important for manager to make decisions because it provides fundamental information to regulate forest. It provides key information for manager to decide the directions of national policies. They need to know how much forest growing stock so that they can have some reference by following the permanent plots for surveying and acquiring the information from data after the processing and analyzing. The large scale forest inventories have its necessities because the small scale surveys are lack of representation. Managers need the large scale inventories to fit the large scale information to formulate the policies. However, using the double sampling methods is the trend worldwide with aerial plots and ground plots. For forest management, not only the growing stock information but also we needs tree location in permanent plots level. Nowadays, most countries around the world use double sampling to conduct the estimating model between air plots and ground plots. In this study, we use ground plots to estimate the forest stocking, multiple regression analysis and K nearest neighbor method to estimate forest stocking. K nearest neighbor method assumed the plots data distribute in a space. Then using different K to decide how many nearest neighbor plots should be choose to calculate the average known values to define the prediction value of unknown point. As the result, this thesis found that ground plots data can provide parameters to map the plots absolute location. And we use ground data to estimate forest stocking is 36,531,244 m3. By stratified sampling method can increase the accuracies, we calculate forest stocking by stratify forest types (33,956,727 m3). Integrating the spatial remote sensing images and inventories data by multiple regression method can estimate forest stocking accurately. After testing the accuracy of different indexes, this study displayed that 3 indexes are better than others, including DVI, RVI and GRVI. This study also found that the accuracy of stratified sampling is higher than system sampling method with many kinds of varieties of forest. Multiple regression estimation models perform good estimating accuracies in different forest types. K nearest neighbor displayed that combining multiple variables can have lower K value and RMSE by topographic variables (elevation, slope and aspect). Using single variable also can estimate the forest stocking but it worse than multiple variables by K value least 14 and RMSE 120 m3/ha. Through adding variables can decrease K value to 6 also reduce RMSE to 101.02 m3/ha. Above all, remote sensing estimation model has fewer RMSE (61.71 m3/ha) than K-nearest neighbor estimation model in the study. Thus remote sensing model has better estimating results. However, with less plots data, multiple regression estimation models show better accuracies at forest stocking estimation. If the estimation should be extended to wider scale, with more plots number, we can use K nearest neighbor method to achieve high accuracy estimating.