论文部分内容阅读
物种生境模型预测结果通常是概率性的,然而在具体的保护管理等实践应用过程中通常需要基于二元值(存在/不存在)的分布图,此时就需要把概率性的预测结果转化为二元值,在此转化过程中就涉及阈值选择问题。此外,在评估模型预测准确度的时候,多数评估指标也需要选择一个阈值用于转化概率预测结果,这个阈值选择对于模型预测准确度也会有极大的影响。然而阈值选择却是物种生境模拟不确定性研究中较少涉及的领域。“随机森林”既可以生成物种生境概率分布图(回归算法)也可以生成二元分布图(分类算法),然而还未见对两种预测方式的比较研究。该文以珙桐(Davidia involucrata)和杉木(Cunninghamia lanceolata)为例,分别采用“随机森林”的分类算法和回归算法预测其生境二元分布图和概率分布图,通过4个不同阈值选择方法(默认值0.5、MaxKappa、MaxTSS和MaxACC)把概率预测图转换为二元分布图,进而比较分析转换结果对模型预估的影响。珙桐不同阈值选择方法所确立的阈值之间存在显著差异,而杉木没有显著差异;两物种模型准确度之间没有显著差异;在预测两物种未来气候条件下的生境面积变化、生境分布区迁移方向和距离以及最适宜海拔分布高度变化时,二元值转换后的回归算法与分类算法之间存在显著差异,但回归算法中各阈值选择方法之间没有显著差异。空间生境分布图的相似性分析表明MaxKappa和MaxTSS法具有最大相似性,分类算法与4种阈值选择方法之间具有最大差异。
Species Habitat Prediction results are usually probabilistic, however, a dichotomous (presence / nonexistence) distribution is usually required for practical applications such as conservation management. In this case, probabilistic prediction results need to be transformed into Binary values, in the conversion process involves the threshold selection problem. In addition, when assessing the accuracy of model predictions, most of the assessment indicators also need to select a threshold for transforming the probability prediction results. This threshold selection will also have a great impact on the model prediction accuracy. However, the threshold selection is one of the less involved fields in the study of modeling uncertainty of species habitat. “Random Forest” can generate the probability distribution of species habitat (regression algorithm) can also generate a binary distribution map (classification algorithm), but no comparison of the two prediction methods. Taking Davidia involucrata and Cunninghamia lanceolata as examples, we use the “random forest” classification algorithm and regression algorithm respectively to predict the habitat binary distribution and probability distribution map. Through four different threshold selection Methods (default values of 0.5, MaxKappa, MaxTSS, and MaxACC) convert probabilistic predictions to binary maps, and then compare the impact of conversion results on the model’s predictions. There was a significant difference between the thresholds established by the different threshold selection methods of Davidia involucrata, but there was no significant difference between the two Chinese fir species. There was no significant difference between the two species models. In predicting the change of the habitat area and the habitat distribution of the two species in the future climatic conditions, There is a significant difference between the regression algorithm and the classification algorithm after the binary value conversion, but there is no significant difference between the threshold selection methods in the regression algorithm. The similarity analysis of spatial Habitat distribution shows that MaxKappa and MaxTSS have the maximum similarity, and the classification algorithm has the biggest difference with the four threshold selection methods.