評估兩種取樣設計對於香桂適生育地模式預測能力之影響

The Effects of Different Sampling Designs on the Ability of Model for Predicting the Suitable Habitat of Cinnamomum subavenium

香桂（Cinnamomum subavenium Miq.）為常綠闊葉樹種，廣泛分布於台灣中、南部山區。本研究矩形試區位於台灣中部，涵蓋惠蓀林場，香桂為此試區的優勢樹種之一，故選為研究的對象。研究目標係藉GIS疊合GPS定位之香桂樣株圖層與海拔、坡度、坡向、坡面位置及SPOT-5影像導出植生指標圖層，協同多變量統計模擬試區香桂之空間分布型態。研究建立邏輯思複迴歸（LMR）、抉擇樹（DT）及區別分析（DA）三種模式，預測並繪製全區的香桂適生育地。建模與驗模採兩種取樣設計，兩者建模樣本相同，取自東峰溪流域，惟驗模樣本分別取自東峰溪與關刀溪兩流域。準確度評估顯示，DT優於LMR，而前二者又遠優於DA。三者於建模、驗模與繪圖之執行效率相當。重要的是DT和LMR於首次模擬，大幅縮小實地調查面積，節省可觀的經費及人力，故兩者更適用於香桂適生育地之模擬。SPOT-5影像導出植生指標改善模式預測能力效用很小，乃因其光譜及空間解析度皆不足，無法分辨散生香桂。三種統計法建立「東峰模式」雖都通過東峰驗模組檢測，但皆未通過關刀驗模組檢測，凸顯此模式無法僅透過地形變數跨越空間精確外推無建模樣本區域。未來研究將嘗試從高空間、高光譜解析度遙測資料萃取物種光譜資訊作建模用變數，期能跨越空間精確外推無建模樣本區域。 Randaishan cinnamons (Cinnamomum subavenium Miq.), one of the evergreen broad-leaved tree species, are generally distributed in central and southern Taiwan. The species was chosen as target for this study because it is one of the dominant species in the Huisun study area in central Taiwan. GIS technique was applied to overlay the tree samples positioned by GPS on the layers of elevation, slope, aspect, terrain position, and vegetation indices derived from SPOT-5 images for modeling the tree's suitable habitat. Decision tree (DT), discriminant analysis (DA), and logistic multiple regression (LMR) models were developed to predict and map the tree's suitable sites in the study area, and to determine the optimum one in terms of accuracy and efficiency. Two sampling designs were created for model development and validation. They used the same set of training samples from Tong-Feng watershed for model development but different sets of test samples for model validation, one from Tong-Feng and the other from Guan-Dau watershed. Accuracy assessment showed that the accuracy of DT was slightly better than that of LMR, accuracies of the two models were much better than that of DA; and the three models were highly efficient in implementation of model development and validation. More importantly, DT and LMR can be applied to predict the tree's suitable habitat because they greatly reduced the area of field survey to 4-7 % of the entire study area at the first stage. Vegetation indices derived from SPOT-5 images could not improve the predicting ability of models for the widely distributed species because of SPOT imagery lacking fine spectral resolution and spatial resolution. The “Tong-Feng models” developed from three methods failed to pass validation by Guan-Dau test samples despite passing validation by Tong-Feng test samples. The outcome emphasized that the “Tong-Feng models” only based on topographic variables could not perform spatial extrapolation accurately from a smaller area with training data to a larger area without any training data. Follow-up studies will attempt to extract spectral information associated with the species from high spatial, spectral resolution remotely sensed data and use it as variable for model development so that the ability of spatial extrapolation with a model can be improved.