應用GIS和邏輯思複迴歸於卡氏櫧潛在生育地之推測

Application of GIS and Logistic Multiple Regression (LMR) to Predict the Potential Habitat of Castanopsis carlesii

卡氏櫧種子為動物的重要食物來源之一，故其在生態體系上具顯著之意義和價值，似更甚於過往所具之經濟價值。大多數研究將地理資訊系統（Geographic Information System, GIS）結合統計應用在瀕危珍稀動植物生育地之模擬，反而少有用在廣泛分布的植物上，所以本研究選擇卡氏櫧為探討對象，以瞭解GIS方法的適用性。本研究以GIS與統計整合分析卡氏櫧於海拔、坡度、坡向及坡面位置等四項地文因子之空間分布特性，並評估四因子對其生育地之相對重要性。研究目標係以這四因子建立邏輯思複迴歸（Logistic Multiple Regression, LMR）、等權、非等權三種模式，用以推測惠蓀林場試區卡氏櫧之潛在生育地，並比較三者的推測準確度與執行效率，從而決定最佳模式。準確度評估結果顯示三者準確度皆在89%以上，其中LMR模式最優，非等權居次，而等權殿後。LMR模式的效率遠優於等權和非等權，而後兩者皆耗時費力。若僅就準確度評估結果來看，三模式皆可用在廣泛型分布卡氏櫧潛在生育地之推測，惟因驗模與建模樣本都來自同一區位難免有空間自相關影響之顧慮，加上驗模組樣本數尚非充足，故此論點仍待再驗證。因此，後續研究將使用來自不同區位之獨立樣本執行驗證，以確認模式之可靠性。 Long-leaf chinkapin (Castanopsis carlesii) has its significance and value in the ecological system, and its value in ecology seems more important than its economic value in the past because the seeds of the tree species are one of the important food sources for the animal. Most studies have applied GIS combined with statistical techniques to model the habitat of the endangered rare species of either plants or animals, whereas a limited number of studies have done the same work on the common, widely distributed tree species. Therefore, long-leaf chinkapin was chosen as a target for the study to understand the applicability of the GIS methods. The study examined the spatial distribution characteristics of the species on the four topographic factors, including elevation, slope, aspect, and terrain position, as well as assessed their relative importance to the chinkapin’s habitat. The research objective was to build logistic multiple regression (LMR) model, equal weight model, and unequal weight model for predicting the potential habitat of the species in the Huisun study area, and to determine the best one in terms of accuracy and efficiency. Accuracy assessment results indicated that the accuracies of the three models are over 89%, and the LMR model is the best one among them and equal weight model is the last. Furthermore, the LMR is the most efficient in implementation, and the other two models are difficult to implement and thus are labor-intensive, time-consuming. Hence, the three models are suitable for predicting the potential habitat of widely distributed chinkapins in terms of model’s accuracy. However, the conclusion needs to be reconfirmed because of spatial autocorrelation probably existing between the sample sets chosen from the same area and insufficient samples for model validation. Thus model validation shall be done with independent samples chosen from the different plots far away from the sample plots for model development in the same study area in follow-up studies so that the reliability of model can be confirmed.

Long-leaf chinkapin (Castanopsis carlesii) has its significance and value in the ecological system, and its value in ecology seems more important than its economic value in the past because the seeds of the tree species are one of the important food sources for the animal. Most studies have applied GIS combined with statistical techniques to model the habitat of the endangered rare species of either plants or animals, whereas a limited number of studies have done the same work on the common, widely distributed tree species. Therefore, long-leaf chinkapin was chosen as a target for the study to understand the applicability of the GIS methods. The study examined the spatial distribution characteristics of the species on the four topographic factors, including elevation, slope, aspect, and terrain position, as well as assessed their relative importance to the chinkapin’s habitat. The research objective was to build logistic multiple regression (LMR) model, equal weight model, and unequal weight model for predicting the potential habitat of the species in the Huisun study area, and to determine the best one in terms of accuracy and efficiency. Accuracy assessment results indicated that the accuracies of the three models are over 89%, and the LMR model is the best one among them and equal weight model is the last. Furthermore, the LMR is the most efficient in implementation, and the other two models are difficult to implement and thus are labor-intensive, time-consuming. Hence, the three models are suitable for predicting the potential habitat of widely distributed chinkapins in terms of model’s accuracy. However, the conclusion needs to be reconfirmed because of spatial autocorrelation probably existing between the sample sets chosen from the same area and insufficient samples for model validation. Thus model validation shall be done with independent samples chosen from the different plots far away from the sample plots for model development in the same study area in follow-up studies so that the reliability of model can be confirmed.