基于Stacking模型集成算法的莲都区南方红豆杉潜在分布区

研究使用R软件中的CaretEnsemble和Caret程序包，并基于Stacking方法来实现模型集成，研究南方红豆杉Taxus chinensis var.mairei在浙江省丽水市莲都区的潜在分布区，并比较5种单一模型的模拟结果及其与集成模型的差异。结果表明:单一模型中极端梯度上升模型表现最好，其次是随机森林模型、支持向量机模型、朴素贝叶斯模型和分类回归树模型，集成模型模拟结果好于单一模型，其Kappa值达0.80，准确率达0.90。集成模型模拟结果显示:影响南方红豆杉分布的主要环境因子为海拔、归一化植被指数和年平均最少降雨量。南方红豆杉主要适宜生长在浙江省丽水市莲都区的山地丘陵地区，中部盆地及平原地区不适宜南方红豆杉的生长，其在莲都区的潜在分布区面积为5.01万hm2。构建的集成模型在一定程度上提高了模型精度，使预测效果更优。

Potential distribution area of Taxus chinensis var. mairei in Liandu District based on a Stacking algorithm

To study the potential distribution of Taxus chinensis var. mairei in Liandu District, the Caret and Caretensemble package in R were used to obtain an ensemble model based on the Stacking method. Then simulation results of five single models[the Extreme Gradient Boosting (XGBoost) Model, the Random Forest (RF) Model, the Support Vector Machine (SVM) Model, the Native Bayes (NB) Model, and the Classification and Regression Tree (CART) Model)] and their differences with the ensemble model were compared. Using 40 presence-only points and generate the same number of pseudo-absences points for modeling, divide the dataset using 10-fold cross-validation and verify model accuracy using Kappa and overall accuracy. Results showed that XGBoost performed best as a single model followed by RF, SVM, NB, and CART. However, the ensemble model was better than all single models with its Kappa value reaching 0.80 and having an overall accuracy of 0.90. According to simulation results of the ensemble model, the main environmental factors affecting the distribution of T. chinensis var. mairei were altitude, normalized difference vegetation index (NDVI), and average annual minimum rainfall. T. chinensis var. mairei was suitable for growing in the mountainous and hilly areas of Liandu District but not in the Central Basin and plains area with the potential area for distribution in Liandu District being 5.01×104 hm2. Overall, the ensemble model used here improved the precision of the model somewhat making the prediction results better.