1. Plant species can show considerable morphological and functional variation along environmental gradients. This intraspecific trait variation (ITV) can have important consequences for community assembly, biotic interactions, ecosystem functions and responses to global change. However, directly measuring ITV across many species and wide geographic areas is often infeasible. Thus, a method to predict spatial variation in a species’ functional traits could be valuable.
2. We measured specific leaf area (SLA), height and leaf area (LA) of grasses across California, covering 59 species at 230 sampling locations. We asked how these traits change along climate gradients within each species and used machine learning to predict local trait values for any species at any location based on phylogenetic position, local climate and that species’ mean traits. We then examined how much these local predictions alter patterns of assemblage-level trait variation across the state.
3. Most species exhibited higher SLA and grew taller at higher temperatures and produced larger leaves in drier conditions. The random forests predicted spatial variation in functional traits very accurately, with correlations up to 0.97. Because trait records were spatially biased towards warmer areas, and these areas tend to have higher SLA individuals within each species, species means of SLA were upwardly biased. As a result, using species means over-estimates SLA in the cooler regions of the state. Our results also suggest that height may be substantially under-predicted in the warmest areas.
4. Synthesis. Using only species mean traits to characterize the functional composition of communities risks introducing substantial error into trait-based estimates of ecosystem properties including decomposition rates or NPP. The high performance of random forests in predicting local trait values provides a way forward for estimating high-resolution patterns of ITV without a massive data collection effort.

中文翻译：

---

预测加利福尼亚草的种内性状变异

1. 植物物种可以沿着环境梯度表现出相当大的形态和功能变化。这种种内性状变异 (ITV) 会对群落组装、生物相互作用、生态系统功能和对全球变化的反应产生重要影响。然而，直接测量许多物种和广泛地理区域的 ITV 通常是不可行的。因此，一种预测物种功能性状空间变异的方法可能很有价值。
2. 我们测量了整个加利福尼亚草的比叶面积 (SLA)、高度和叶面积 (LA)，覆盖了 230 个采样点的 59 个物种。我们询问了这些特征如何沿着每个物种内的气候梯度变化，并使用机器学习根据系统发育位置、当地气候和该物种的平均特征来预测任何地点任何物种的当地特征值。然后，我们检查了这些局部预测在多大程度上改变了整个州的组合水平特征变化模式。
3. 大多数物种表现出更高的 SLA，在更高的温度下长得更高，并在干燥的条件下产生更大的叶子。随机森林非常准确地预测了功能性状的空间变异，相关性高达 0.97。由于性状记录在空间上偏向于较温暖的地区，并且这些地区在每个物种中往往具有较高的 SLA 个体，因此 SLA 的物种均值向上偏向。因此，使用物种意味着高估了该州较冷地区的 SLA。我们的结果还表明，在最温暖的地区，高度可能被大大低估。
4. 合成。仅使用物种平均特征来表征群落的功能组成，可能会给基于特征的生态系统属性估计引入大量误差，包括分解率或 NPP。随机森林在预测局部特征值方面的高性能提供了一种无需大量数据收集工作即可估计 ITV 高分辨率模式的方法。