基于机器学习方法的景观斑块属性对物种丰富度的影响研究

Influence of landscape patch properties on species abundance based on machine learning method

景观破碎化对物种的影响依赖于斑块属性,斑块属性体现了斑块的类型、面积、形状及其与周围斑块间关系。然而,在破碎化农业景观中这些斑块属性对关键物种的联合影响仍知之甚少。选取中国海南地区的典型农业破碎化景观,实测了180个斑块中三类关键物种,分别为害虫-橡胶材小蠹虫(Xyleborus affinis)、传粉昆虫-中华蜜蜂(Apis cerana)和天敌物种-玉米螟赤眼蜂(Trichogrammatid ostriniae Pang et Chen),刻画了各斑块属性,包括斑块类型、斑块面积、形状指数、距天然林距离、是否与橡胶接触、景观中天然林所占比例、是否有林下植物,并使用3种机器学习模型(神经网络、随机森林、支持向量机)拟合斑块属性与三类物种丰富度关系,解析斑块属性的相对重要性。结果表明,与支持向量机和神经网络相比,随机森林模型对三个物种的丰富度预测效果均最好,其R2在橡胶材小蠹虫、中华蜜蜂和赤眼蜂中分别达到0.785、0.845和0.798;基于随机森林模型的结果表明,斑块类型对橡胶材小蠹虫和中华蜜蜂丰富度相对重要性高于其他斑块属性;斑块面积对赤眼蜂丰富度影响高于其他斑块属性,且与斑块类型存在交互作用;橡胶林中橡胶材小蠹虫与中华蜜蜂丰富度高于天然林和其他斑块类型,玉米螟赤眼蜂丰富度则在天然林与农田中最高。研究揭示了斑块属性对不同物种丰富度的相对影响,阐明了斑块类型在不同物种中的重要作用,并证实了机器学习是推导斑块属性对物种影响的有效方法。

The influence exerted by landscape fragmentation on various species is intricately tied to the properties of individual patches,which encompass a multitude of factors such as the type of patch,its area,shape,and the dynamic interactions these patches maintain with their neighboring counterparts.Within the specific context of fragmented agricultural landscapes,the exact impact of these patch properties on key species is a domain that still remains shrouded in a veil of limited understanding and empirical clarity.This study pivots its focus towards a representative agricultural landscape located in the Hainan region of China,a landscape that is notably marked by substantial fragmentation.In this particular context,a meticulously conducted field investigation spanned across 180 patches,each serving as a distinct habitat for three critical species:pests(Xyleborus affinis),pollinators(Apis cerana),and natural enemies(Trichogrammatid ostriniae Pang et Chen).The properties of each patch,encompassing elements such as type,area,shape,and the intricate interplay with adjacent patches,were rigorously documented and analyzed.Subsequently,to decipher the complex relationships between the properties of these patches and the abundance of the aforementioned species,the capabilities of three advanced machine learning models were harnessed:namely,the Artificial Neural Network,the Random Forest,and the Support Vector Regression models.These models were adeptly employed,leveraging their respective analytical strengths,to analyze the intricate relationships at play.The results of this comprehensive study uncovered a significant revelation:the robust Random Forest model emerged as a superior predictor of the abundance of all three species in comparison to the Support Vector Regression and the Artificial Neural Network models.This was evidenced by impressively high R-squared values:0.785 for X.affinis,0.845 for A.cerana,and 0.798 for T.ostriniae.These findings underscored the paramount importance of certain patch properties,most notably the type of patch,in influencing the abundance of these species.In particular,it was observed that patch area exerted a dominant influence on the abundance of T.ostriniae,with subtle yet crucial interactions influencing the abundance of X.affinis and A.cerana,thereby surpassing the significance of other patch properties.Furthermore,variations in patch types revealed intriguing patterns.Notably,the study unveiled that rubber plantations harbored flourishing communities of X.affinis and A.cerana,a finding that was markedly pronounced when contrasted with the abundance found in natural forests and other patch categories.In stark contrast,T.ostriniae exhibited its highest abundance in the pristine sanctuaries of natural forests and cultivated landscapes.Overall,this study not only reveals the differential impacts of patch properties on the richness of multiple species but also elucidates the critical role that specific patch types play for different species.Additionally,it validates the efficacy of machine learning as a potent and insightful tool for inferring the impacts of landscape fragmentation and patch properties on species abundance and diversity.