應用3S技術及多變量統計於薯豆及卡氏櫧潛在生育地之推估

Application of 3S and Multivariate Statistics to Predict the Potential Habitat of Elaeocarpus Japonicus and Castanopsis carlesii

薯豆（Elaeocarpus japonicus Sieb. et Zucc.）和卡氏櫧（Castanopsis carlesii(Hemsl.) Hay.）為常綠闊葉樹種，廣泛分布於台灣山區。本研究矩形試區位於台灣中部，涵蓋惠蓀林場，因薯豆為此試區的優勢樹種之一，且因卡氏櫧種子為此試區動物的重要食物來源之一，故選兩者為研究的對象。本研究目標係應用3S地球空間資訊技術結合多變量統計模擬並測繪惠蓀林場薯豆及卡氏櫧之潛在生育地。本研究以GIS將取自東峰溪與關刀溪流域兩樹種樣木GPS定位圖層與海拔、坡度、坡向、坡面位置、SPOT-5影像導出植生指標等五圖層加以疊合，分析試區二樹種之空間分布型態，從而建立抉擇樹（DT）、邏輯思複迴歸（LMR）及區別分析（DA）三種模式，推測並繪製兩樹種的潛在生育地。結果顯示，此二樹種於DT模式推測準確度略優於LMR者，而它們又遠高於DA；三者建模及驗模效率大致旗鼓相當，惟DT在全區潛在圖測繪略嫌繁瑣。更重要的是三模式於首次推測，即由全區篩選3-12%兩樹種的高潛力區，大幅縮減後續地面調查之面積，節省龐大人力與經費。最重要的是，卡氏櫧三種模式推測精度分別高於薯豆相對應三模式，從生態學角度來看，乃因卡氏櫧限制因子為海拔1, 700m，而薯豆卻無此限制因子。因此，雖同屬廣泛分布型樹種，卡氏櫧在物種空間分布模擬相對上較薯豆容易，且模擬成果較佳。 Japanese Elaeocarpus (Elaeocarpus japonicas, JE) and long-leaf chinkapins (Castanopsis carlesii, LCC) evergreen broad-leaved tree species are widely distributed in Taiwan. The two species were chosen as target for this study because the former is one of the dominant species and because the seeds of the latter are one of the important food sources for the animal in the Huisun study area. The study objective was to apply 3S geo-spatial information technologies, including GIS, RS, and GPS, to predict the potential habitat of the two tree species in the study area. We overlaid the layers of the tree samples collected with GPS on the layers of elevation, slope, aspect, terrain position, and vegetation indices derived from SPOT-5 images for analyzing the spatial pattern of the two tree species by using a GIS. We developed the decision tree (DT), logistic multiple regression (LMR) and discriminant analysis (DA) models for modeling the distribution of the two tree species. Accuracy assessments based on the tree samples of the two species collected from Tong-Feng and Kuan-Dau watershed in Huisun indicated that the accuracies of DT were somewhat better than those of LMR, and both were much better than those of DA. These three models were very efficient in model development and validation, but DT was somewhat difficult to map the potential habitat for the entire area. More importantly, the three models greatly reduced the area of field survey to 3-12 % of the entire area at the first stage of modeling, thus saving much labor and cost for subsequent field survey. Most importantly, the accuracies of the three models with LCC were much better than those with JE, respectively, because elevation at 1,700 m is a limiting factor of LCC and JE does not have such a limiting factor from the ecological standpoint. Consequently, modeling spatial distribution for LCC was much easier and better than that for JE, even though LCC and JE are widely distributed in Taiwan.

Japanese Elaeocarpus (Elaeocarpus japonicas, JE) and long-leaf chinkapins (Castanopsis carlesii, LCC) evergreen broad-leaved tree species are widely distributed in Taiwan. The two species were chosen as target for this study because the former is one of the dominant species and because the seeds of the latter are one of the important food sources for the animal in the Huisun study area. The study objective was to apply 3S geo-spatial information technologies, including GIS, RS, and GPS, to predict the potential habitat of the two tree species in the study area. We overlaid the layers of the tree samples collected with GPS on the layers of elevation, slope, aspect, terrain position, and vegetation indices derived from SPOT-5 images for analyzing the spatial pattern of the two tree species by using a GIS. We developed the decision tree (DT), logistic multiple regression (LMR) and discriminant analysis (DA) models for modeling the distribution of the two tree species. Accuracy assessments based on the tree samples of the two species collected from Tong-Feng and Kuan-Dau watershed in Huisun indicated that the accuracies of DT were somewhat better than those of LMR, and both were much better than those of DA. These three models were very efficient in model development and validation, but DT was somewhat difficult to map the potential habitat for the entire area. More importantly, the three models greatly reduced the area of field survey to 3-12 % of the entire area at the first stage of modeling, thus saving much labor and cost for subsequent field survey. Most importantly, the accuracies of the three models with LCC were much better than those with JE, respectively, because elevation at 1,700 m is a limiting factor of LCC and JE does not have such a limiting factor from the ecological standpoint. Consequently, modeling spatial distribution for LCC was much easier and better than that for JE, even though LCC and JE are widely distributed in Taiwan.