Scholarlink

<p>In this perspective, I argue that reef science has the potential to overcome the conventional limits set by scientific explanatory variable selectivity and case studies. Emerging formal data sharing, collaboration, satellites and shipboard data, fine-scale mapping, and human demographic data now eclipse and better identify threats than accumulating and comparing local case studies. Moreover, predictive algorithms combined with geographic information systems are increasingly able to spatially evaluate big data using multivariate associative predictions better than common and variable-limited explanatory models. As an example, I present a summary of 15 journal articles that have used machine learning (ML) to evaluate reef threats on large national or provincial scales and find they have similarly high model fits (<i>r</i><sup>2</sup> = 0.40–0.80), have clear and often similar environmental rankings, no single variable has high importance, all best models require modifying variables, variable selection processes affect interpretations, and predictions are less catastrophic than variable-limited explanatory models. I illustrate these findings with a synthesis of common reef status metrics (coral cover, fish biomass, and numbers of taxa) for an evaluation of the Western Indian Ocean Province (WIO) using a Boosted Regression Tree analysis. Predictive algorithms identify the dominant role of modifying variables, which are primarily acute and chronic temperature variability. The dominant explanatory climate change variables of excess temperature and ocean acidity are shown to be either weak or inferring a directional response opposite to the explanatory climate model predictions. Until explanatory and predictive model relationships coincide, future predictions for coral reefs will be inaccurate.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Evaluating coral reef hazards requires both explanatory and predictive models

  • 期刊论文

  • 发表日期：

    2025年10月14日

查看全文
  • T. R. McClanahan

摘要

In this perspective, I argue that reef science has the potential to overcome the conventional limits set by scientific explanatory variable selectivity and case studies. Emerging formal data sharing, collaboration, satellites and shipboard data, fine-scale mapping, and human demographic data now eclipse and better identify threats than accumulating and comparing local case studies. Moreover, predictive algorithms combined with geographic information systems are increasingly able to spatially evaluate big data using multivariate associative predictions better than common and variable-limited explanatory models. As an example, I present a summary of 15 journal articles that have used machine learning (ML) to evaluate reef threats on large national or provincial scales and find they have similarly high model fits (r2 = 0.40–0.80), have clear and often similar environmental rankings, no single variable has high importance, all best models require modifying variables, variable selection processes affect interpretations, and predictions are less catastrophic than variable-limited explanatory models. I illustrate these findings with a synthesis of common reef status metrics (coral cover, fish biomass, and numbers of taxa) for an evaluation of the Western Indian Ocean Province (WIO) using a Boosted Regression Tree analysis. Predictive algorithms identify the dominant role of modifying variables, which are primarily acute and chronic temperature variability. The dominant explanatory climate change variables of excess temperature and ocean acidity are shown to be either weak or inferring a directional response opposite to the explanatory climate model predictions. Until explanatory and predictive model relationships coincide, future predictions for coral reefs will be inaccurate.