運用背景資料取樣策略增進物種分布模型的預測效能

Improving the predictive performance of species distribution models using background data sampling strategies

物種分布模型中，最大熵（maximum entropy, MaxEnt）的效能高度依賴輸入數據的質量，特別是物種出現紀錄的代表性，然而，實際研究中常因取樣偏差導致模型準確性和泛化能力下降。由於背景資料的選擇對於校正取樣偏差至關重要，本研究旨在比較不同背景資料取樣方法對MaxEnt模型性能的影響，並以臺灣特有且易危的霧社櫻（Prunus taiwaniana）為研究對象，探討其分布預測的最佳策略。總共比較了五種背景資料取樣方法，包括目標群體密度、目標種密度、隨機取樣、受限距離及最小多邊範圍。結果顯示，目標群體密度和目標種密度方法表現佳，其Boyce指數都達到0.9以上，表示這兩種方法能有效校正取樣偏差並捕捉廣泛的環境條件。此外，從最佳模型的預測結果可知，霧社櫻偏好寒冷環境，並對特定的降水量與晝夜溫差等氣候條件具有高度敏感性，其高適宜棲地主要集中在臺灣中北部的中海拔山區。本研究強調合理選擇背景資料取樣策略的重要性，特別是對於稀有或受脅物種，適當的取樣方法不僅能顯著提升模型的準確性，還能為保育工作提供重要的科學依據。

In species distribution models, the performance of maximum entropy（MaxEnt）highly depends on the quality of the input data, particularly the representativeness of species occurrence records. However, sampling bias in practical studies often leads to reduced model accuracy and generalization capabilities. Because the choice of background data is crucial for correcting sampling bias, this study aimed to compare the effects of different background data sampling methods on MaxEnt model performance, using the endemic and vulnerable plant species（Prunus taiwaniana）as the research subject to explore the best strategy for predicting its distribution. Five background data sampling methods were compared: target group density, target species density, random sampling, constrained distance, and minimum convex polygon. The results showed that target group density and target species density performed well, with Boyce index values exceeding 0.9, indicating that these two methods effectively correct sampling bias and capture a wide range of environmental conditions. Furthermore, predictions from the best-performing model revealed that P. taiwaniana prefers cold environments and exhibits high sensitivity to specific climatic conditions, such as precipitation levels and diurnal temperature variation. The most suitable habitats are primarily concentrated in the mid-elevation mountainous regions of central and northern Taiwan. This study emphasizes the importance of selecting appropriate background data sampling strategies, particularly for rare or threatened species. Proper sampling methods not only significantly improve model accuracy but also provide crucial scientific support for conservation efforts.