The demography of rare species: building for whole forest prediction and assessment/ La démographie des espèces rares : construire une prévision et une évaluation de la forêt entière

Petit séjour d’étude organisé par Sean McMahon, du 21 au 31 mai 2024

Participants

David BAUMAN (Institut de Recherche pour le Développement – Délégation Régionale Occitanie – Montpellier, France), Claire FORTUNEL (IRD, UMR AMAP – Montpellier, France), Sean MCMAHON (Smithsonian Institution – Edgewater, États-Unis)

Résumé

Forest on Barro Colorado

Les espèces rares constituent la majorité des arbres des forêts tempérées et tropicales et jouent un rôle particulier dans ces forêts. Pourtant, en raison du manque évident de données, les espèces rares sont mal représentées ou absentes des modèles à grande échelle de la dynamique forestière. Notre atelier comprenait trois scientifiques qui se sont concentrés sur les aspects biologiques, écologiques et quantitatifs de l’intégration des espèces rares dans les évaluations de la fonction forestière. Nous avons recueilli des données, préparé des données pour l’analyse et commencé des analyses au cours de notre atelier de dix jours, en nous concentrant sur les données d’une forêt tropicale au Panama. Pour chaque étape du processus et chaque décision prise, nous avons utilisé un flux de travail codage-texte. Il en résulte une liste de documents (pdf et html) qui guide la préparation des données et les étapes d’analyse tout en les expliquant, ce qui aboutit à un produit qui peut à la fois éduquer d’autres scientifiques sur la manière dont nous avons abordé ce projet et promouvoir un exemple de science transparente et reproductible. À la fin de l’atelier, nous avions lancé des analyses à l’aide d’une approche d’apprentissage automatique qui incluait 284 espèces allant d’une seule à 20 000 observations, y compris plus de 170 variables de traits (“caractéristiques”) allant des propriétés chimiques et physiques des feuilles à la physique du bois et aux stratégies démographiques. Nous avons d’abord étudié ce qui “fait” une espèce rare en utilisant des simulations de petites populations échantillonnées à partir de grandes populations “connues”. Nous avons ensuite formé un modèle de régression par apprentissage automatique sur les taux de croissance maximaux de 120 espèces communes afin de prédire la fonction de 164 espèces rares. Bien qu’il reste des défis importants à relever, nous avons fait des progrès étonnants, tant dans la conceptualisation que dans la mise en œuvre de ce programme important mais difficile.

Summary

Numerically rare species constitute the majority of trees in temperate and tropical forests and serve special roles in those forests. Yet, due to the obvious lack of data, rare species are poorly represented or absent from large scale models of forest dynamics. Our workshop included three scientists focused on the biological, ecological, and quantitative aspects of incorporating rare species into assessments of forest function. We gathered data, prepared data for analysis, and began analyses during our ten-day workshop, focusing on data from a tropical forest in Panama. Critically, for every step of the process and every decision made we employed a coding-text workflow. This resulted in a document pipeline (both pdf and html) that guides data preparation and analytical steps while explaining them, leading to a product that can both educate other scientists in how we approached this project and advance an example of transparent, reproducible science. By the end of the workshop we had initiated analyses using a machine learning approach that included 284 species ranging from a single to 20 thousand observations, including over 170 trait variables (‘features’) ranging from chemical and physical properties of leaves to wood physics and demographic strategies. We first explored what ‘makes’ a rare species using simulations of small populations sampled from ‘known’ large populations. We then trained a machine learning regression model on the maximum growth rates of 120 common species in order to predict the function of 164 rare species. Although significant challenges remain, we made astounding progress, both in the conceptualization and implementation of this important but challenging program.

Report

Rare species form an integral part of forest communities, contributing in critical ways to forest structure, diversity and dynamics. For example, emergent tree species (species that rise above the height of the regular canopy) are numerically rare and yet contain a large fraction of the carbon in a tropical forest. We can quantify little about the drivers of their growth and death rates because there are so few stems compared to the more common canopy species, leaving our classical statistical approaches empty. Rare species are also at higher risk of extirpation or extinction due to their low numbers and potential niche specialization within the forest. Our working group was designed to take on this great challenge in a way that would map the problem, and develop a coding solution, which could be used for the great wealth of data regularly being collected through forest inventory initiatives.

This resulted in a wonderful ‘back-and-forth’ series of discussions, small and large, between all members of the group. We needed to accomplish practical products and advances, all while checking with each other on the decisions we made, what to include or not, how to format, describe, or render data, and how to design and implement machine learning analyses. Les Treilles was perfect for this, as we could enjoy these discussions not by zoom or a far-flung meeting, but daily, regularly, during meals, on walks, over evening card games. This constant, parallel, and immersive investment into the multiple challenges to the project at hand was indispensable to our advancement of the program.

One way we sought to cement the work we were doing at Les Treilles was by building an analytical ‘pipeline’ that spanned data-fusion modeling, machine learning algorithm application, and ecological inference. This approach ensured that after our working group, the legacy of this intense investment of intellectual and collaborative, creative energy, would have a documented legacy, one we could easily continue after the working group. But it was also one we could easily communicate to our larger community. To build this pipeline we used the Quarto open-source scientific publishing platform, which links text, code, and code implementation to produce documents in pdf and html that upon compiling replicate the entire workflow. This facilitates open and reproducible scientific enquiry.
The day to day practical component of our working group focused on three aspects of the larger program:

1) soliciting, vetting, and prepping demographic data—especially data on the growth and mortality of tropical forest species from Panama;

2) the collection and preparation of ‘feature’ data, including taxonomic data, phylogenetic relationships between data, and functional traits of species (we collected over 170 such features from the literature and personal work);

3) the simulation of rare species data built from common species to identify the ways in which having fewer data points might bias or inhibit estimates of important dynamics of rare species given that we had robust estimates from common species; 

4) and the implementation of a machine algorithm to translate inference from data that showed strong feature and response coverage to those species that were sparse.

We used data from Barro Colorado Island (BCI), a well-studied 50 ha. forest in Panama with over 300K stems identified to species, mapped, and measured every five years (since 1981). These were data with which we were all familiar, and yet required considerable effort to translate into the form we needed to advance our objectives. Of the 284 total species in the plot, 168 have fewer than 100 stems, and 64 have fewer than 10. To put this in perspective, annual tree mortality rates are ~ 2% (1 in 50 individuals might die), a majority of species in a large tropical forest have insufficient samples to apply any statistical assessment of mortality causes and consequences. This was an ideal data set. Further, there is a wealth of potential predictors of the behavior of these species through intensive collections of plant functional traits that have occurred over the past two decades.

The implications of this product are potentially profound. Our pipeline will allow multiple data sources and advanced classification techniques to assign such ‘types’ to entire forests, with uncertainty, and the ability to re-assess those classifications when new information arises. Concepts and results from this work, with attribution to Les Treilles, has already been presented to colleagues in two talks and many email and zoom discussions. We will advance this program over the coming months and will produce important papers (methodological and ecological) born from this work.