NOAA/WDS Paleoclimatology - Gut - Eschenbach-Cholwald - PCAB - ITRDB SWIT369

The pairwise statistical comparison of ring-width series is the basic analysis of dendro-provenancing studies. It is assumed that statistical proximity indicates similar provenance, but this assumption often remains untested. Especially for small areas with high topographic complexity, it is unknown to what extent statistical proximity and geographical provenance are correlated.

In this paper, dendro-provenancing is framed as a search for statistical Nearest Neighbors. The 'k-Nearest Neighbors leave one-out cross-validation' process (k-NN) is proposed as a method for validating dendro-provenancing approaches. Furthermore, it allows researchers to consistently compare and evaluate different proximity measures with respect to their suitability for dendro-provenancing. The validation process is demonstrated on a data set of 401 ring-width series of Norway spruce (Picea abies (L.) H. Karst.) encompassing 15 sites along elevational gradients in north-eastern Switzerland. Moreover, a new type of plot, the so-called scissor plot, is introduced to visualize the k-NN validation process.

Results indicate that dendro-provenancing depends heavily on differences in between sites high-frequency signal. Mean classification success for the relevant stages of the k-NN (CSRopen) ranged from 71.8% to 79.2% for the best performing measures. Classification errors occurred mainly between sites at elevations of 1000-1198 m a.s.l. At all other elevations and between different regions of the study area, only moderate differences in classification performance were detected. Thus, the results indicate that dendro-provenancing may be principally feasible even in a small region as studied here.

树木年轮宽度序列的两两统计比较是树木年轮产地溯源（dendro-provenancing）研究的基础分析手段。现有研究普遍假设，统计相似性越高代表样本的原产地越相近，但这一假设往往未得到实证检验。尤其是在地形复杂的小型研究区域内，统计相似性与地理原产地之间的关联程度仍不明确。

本研究将树木年轮产地溯源转化为统计意义上的最近邻搜索问题。本文提出采用“k近邻留一交叉验证（k-NN）”流程作为验证树木年轮产地溯源方法的通用框架，此外，该流程可使研究者系统性地对比、评估不同相似性度量方法在树木年轮产地溯源任务中的适配性。本研究以瑞士东北部沿海拔梯度分布的15个样点的401条挪威云杉（Picea abies (L.) H. Karst.）年轮宽度序列数据集为对象，演示了该验证流程。此外，本文还提出一种新型可视化图表——剪刀图（scissor plot），用于展示k-NN验证流程的结果。

研究结果表明，树木年轮产地溯源的效果高度依赖于样点间高频信号的差异。针对表现最优的相似性度量方法，k-NN相关阶段（CSRopen）的平均分类准确率介于71.8%至79.2%之间。分类错误主要集中在海拔1000~1198米的样点之间；在其余海拔区间以及研究区内不同区域之间，分类性能仅存在小幅差异。综上，本研究结果表明，即便在本研究这类小型区域内，树木年轮产地溯源在原则上仍具备可行性。