\makeatletter
\@ifundefined{HCode}
{\documentclass[CRBIOL,Unicode,screen,biblatex,published]{cedram}
\addbibresource{crbiol20240549.bib}
\newenvironment{noXML}{}{}
\let\citep\parencite
\let\citet\textcite
\def\xcitealp#1#2{\citeauthor{#1}, \citelink{#1}{#2}}
\newcommand*{\citelink}[2]{\hyperlink{cite.\therefsection @#1}{#2}}
\def\defcitealias#1#2{}
\let\citepalias\parencite
\let\citetalias\textcite
\newenvironment{Table}{\begin{table}}{\end{table}}
\def\thead{\noalign{\relax}\hline}
\def\endthead{\noalign{\relax}\hline}
\def\tabnote#1{\vskip4pt\parbox{.83\linewidth}{#1}}
\def\tsup#1{$^{{#1}}$}
\def\tsub#1{$_{{#1}}$}
\RequirePackage{etoolbox}
\def\jobid{crbiol20240549}
%\graphicspath{{/tmp/\jobid_figs/web/}}
\graphicspath{{./figures/}}
\newcounter{runlevel}
\let\MakeYrStrItalic\relax
\def\refinput#1{}
\def\back#1{}
\def\tbody{\noalign{\relax}}
\newenvironment{inftab}{}{}
\def\hyphen{\text{-}}
\def\xmorerows#1#2{{#2}}
\def\0{\phantom{0}}
\def\xsection#1{}
\def\botline{\\\hline}
\usepackage{multirow}
\def\nrow#1{\@tempcnta #1\relax%
\advance\@tempcnta by 1\relax%
\xdef\lenrow{\the\@tempcnta}}
\def\morerows#1#2{\nrow{#1}\multirow{\lenrow}{*}{#2}}
\DOI{10.5802/crbiol.184}
\datereceived{2024-07-07}
\daterevised{2025-03-03}
\datererevised{2025-04-30}
\dateaccepted{2025-07-07}
\ItHasTeXPublished
\def\og{\guillemotleft}
\def\fg{\guillemotright}
\makeatletter
\g@addto@macro{\UrlBreaks}{\UrlOrds}
\gappto{\UrlBreaks}{\UrlOrds}
\usepackage{hyperref}
\makeatother
}
{
\PassOptionsToPackage{authoryear}{natbib}
\documentclass[crbiol]{article}
\def\CDRdoi{10.5802/crbiol.184}
\let\newline\break
\def\selectlanguage#1{}
\usepackage[T1]{fontenc}
 \def\citelink#1#2{\citeyear{#1}}
 \def\xcitealp#1#2{\citealp{#1}}
\def\href#1#2{\url[#1]{#2}}
\def\xsection#1{}
\makeatletter
\def\CDRsupplementaryTwotypes#1#2{}
}
\makeatother

\usepackage{upgreek}

\dateposted{2025-09-29}
\begin{document}

\begin{noXML}

\CDRsetmeta{articletype}{research-article}

\title{Routine production of population trends from citizen science
data: insights into the dynamics of common bird and plant species in
France}

\alttitle{Produire les tendances de populations en routine \`{a} partir
de suivis participatifs de biodiversit\'{e} : une application aux
esp\`{e}ces communes d'oiseaux et de plantes en France
m\'{e}tropolitaine}

\author{\firstname{Mathilde} \lastname{Vimont}\IsCorresp}
\address{MNHN -- CESCO, 43 rue Buffon, 75005 Paris, France}
\email[M. Vimont]{mathilde.vimont@gmail.com}

\author{\firstname{Lise} \lastname{Bartholus}}
\addressSameAs{1}{MNHN -- CESCO, 43 rue Buffon, 75005 Paris, France}
\email[L. Bartholus]{lise.bartholus@mnhn.fr}

\author{\firstname{Yves} \lastname{Bas}}
\addressSameAs{1}{MNHN -- CESCO, 43 rue Buffon, 75005 Paris, France}
\email[Y. Bas]{yves.bas@mnhn.fr}

\author{\firstname{Beno\^{i}t} \lastname{Fontaine}\CDRorcid{0000-0002-1017-5643}}
\addressSameAs{1}{MNHN -- CESCO, 43 rue Buffon, 75005 Paris, France}
\email[B. Fontaine]{benoit.fontaine@mnhn.fr}

\author{\firstname{Colin} \lastname{Fontaine}\CDRorcid{0000-0001-5367-5675}}
\addressSameAs{1}{MNHN -- CESCO, 43 rue Buffon, 75005 Paris, France}
\email[C. Fontaine]{colin.fontaine@mnhn.fr}

\author{\firstname{Romain} \lastname{Julliard}\CDRorcid{0000-0001-8897-2514}}
\addressSameAs{1}{MNHN -- CESCO, 43 rue Buffon, 75005 Paris, France}
\email[R. Julliard]{romain.julliard@mnhn.fr}

\author{\firstname{Gr\'{e}goire} \lastname{Lo\"{i}s}}
\addressSameAs{1}{MNHN -- CESCO, 43 rue Buffon, 75005 Paris, France}
\email[G. Lo\"{i}s]{gregoire.lois@mnhn.fr}

\author{\firstname{Romain} \lastname{Lorrilli\`{e}re}}
\addressSameAs{1}{MNHN -- CESCO, 43 rue Buffon, 75005 Paris, France}
\email[R. Lorrilli\`{e}re]{romain.lorrilliere@mnhn.fr}

\author{\firstname{Gabrielle} \lastname{Martin}\CDRorcid{0000-0002-9192-4924}}
\address{Paul Sabatier University - Toulouse III - EDB Lab, France} 
\email[G. Martin]{gabrielle.martin@univ-tlse3.fr}

\author{\firstname{Emmanuelle} \lastname{Porcher}\CDRorcid{0000-0002-9264-8239}}
\addressSameAs{1}{MNHN -- CESCO, 43 rue Buffon, 75005 Paris, France}
\email[E. Porcher]{emmanuelle.porcher@mnhn.fr}

\keywords{\kwd{Data analysis pipeline}
\kwd{Citizen-science}
\kwd{Biodiversity monitoring}
\kwd{Population trends}
\kwd{Baseline}
\kwd{Plants}
\kwd{Birds}}

\altkeywords{\kwd{Routine danalyse}
\kwd{Sciences participatives}
\kwd{Esp\`eces communes}
\kwd{Tendances de population}
\kwd{Oiseaux}
\kwd{Plantes}}

\begin{abstract} 
The ongoing environmental crisis, driven by human activities, has
resulted in significant biodiversity losses across various taxa,
affecting ecosystem functioning. To deal with this crisis, policymakers
have notably established the Kunming-Montreal Global Biodiversity
Framework, which includes targets to mitigate biodiversity loss by
2050. To achieve this goal, reliable and ecologically relevant
indicators are essential to quantify and qualify biodiversity changes.
Temporal trends in species abundance or occurrence have been proposed
as useful indicators. In France, the Vigie-Nature program engages
volunteers in biodiversity monitoring through various schemes, thereby
producing relevant data to estimate country-wide temporal trends for
various taxonomic groups. Some indicators of population trends are
already produced for some taxa, but the analysis pipelines remain
unpublished and need extensions to accommodate monitoring schemes
collecting presence/absence instead of abundance data, such as the
Vigie-flore plant monitoring scheme. Here, we present a newly developed
analysis pipeline to estimate population trends, which handles
different data types and protocol specificities, and goes beyond linear
population trends by considering multiple time periods and visualizing
non-linear dynamics. In addition to introducing the methodology and
making it available, we ran this pipeline to produce population trends
for 148 bird and 181 plant species in France, based on abundance data
from STOC (French Breeding Bird Survey) and occurrence data from
Vigie-flore schemes. Results show as many increasing as decreasing bird
population trends over the past 23 years, and a tendency for more
decreasing than increasing plant population trends over the past 15
years, thereby revealing significant changes in community composition.
Specifically, for birds, most habitat generalist species showed stable
or increasing population trends, while most habitat specialist species
showed stable or decreasing population trends, suggesting biotic
homogenization. This pipeline and first analyses provide an
unprecedented overview of bird and plant population trends, and
contribute to the production of biodiversity indicators based on open
science and reproducible research.\looseness=-1
\end{abstract}

\begin{altabstract} 
Les pressions anthropiques \`{a} l'origine de la crise environnementale
actuelle entra\^{i}nent une perte de biodiversit\'{e} pour de nombreux
groupes taxonomiques, qui affecte le fonctionnement des
\'{e}cosyst\`{e}mes. Pour faire face \`{a} cette crise, les
d\'{e}cideurs politiques ont notamment mis en place le Cadre mondial de
Kunming-Montr\'{e}al pour la biodiversit\'{e}, qui comprend des
objectifs visant \`{a} att\'{e}nuer la perte de biodiversit\'{e} d'ici
2050. Pour atteindre ces objectifs, il est essentiel de disposer
d'indicateurs fiables et pertinents sur le plan \'{e}cologique afin de
quantifier et de qualifier les changements en mati\`{e}re de
biodiversit\'{e}. Les tendances temporelles bas\'{e}es sur
l'abondance de populations ou la pr\'{e}sence des esp\`{e}ces
figurent parmi ces indicateurs. Les programmes de sciences
participatives, tels que Vigie-Nature en France, impliquent des
b\'{e}n\'{e}voles non professionnels dans le suivi de la
biodiversit\'{e} et offrent ainsi une solution de r\'{e}colte de
donn\'{e}es \`{a} large \'{e}chelle. Bien que des indicateurs de
tendances de population produits \`{a} partir de donn\'{e}es de
sciences participatives existent pour certains taxons, il n'existe pas
encore d'outil permettant de prendre en compte la diversit\'{e} des
donn\'{e}es issues de ce type de programmes et de caract\'{e}riser la
non-lin\'{e}arit\'{e} des tendances. Nous proposons donc ici une
routine d'analyse flexible, d\'{e}velopp\'{e}e pour estimer les
tendances des populations \`{a} partir de diff\'{e}rents types de
donn\'{e}es de comptage et de pr\'{e}sence-absence. La
m\'{e}thodologie et les codes sont mis \`{a} disposition dans ce
papier, dans lequel nous pr\'{e}sentons \'{e}galement les tendances de
populations pour 148 esp\`{e}ces d'oiseaux et 181 esp\`{e}ces de
plantes en France m\'{e}tropolitaine, \`{a} partir des donn\'{e}es
issues des programmes du STOC (suivi temporel des oiseaux communs, qui
recense les oiseaux nicheurs) et de Vigie-flore. Les r\'{e}sultats
obtenus montrent un comportement diff\'{e}rent des deux taxons. On note
en effet une proportion similaire d'esp\`{e}ces d'oiseaux dont les
tendances de population diminuent ou augmentent, au cours des 23
derni\`{e}res ann\'{e}es. Les esp\`{e}ces de plantes sont davantage
concern\'{e}es par des d\'{e}clins de population que des augmentations.
Dans les deux cas, ces tendances r\'{e}v\`{e}lent des changements
significatifs dans la composition des communaut\'{e}s. Plus
pr\'{e}cis\'{e}ment, pour les oiseaux, la majorit\'{e} des esp\`{e}ces
g\'{e}n\'{e}ralistes montrent des tendances d\'{e}mographiques stables
ou en augmentation, tandis que la plupart des esp\`{e}ces
sp\'{e}cialistes montrent des tendances d\'{e}mographiques stables ou
en diminution, sugg\'{e}rant ainsi une homog\'{e}n\'{e}isation
biotique. Ces conclusions soulignent que l'outil d'analyse propos\'{e}
ici et les premiers r\'{e}sultats obtenus fournissent un aper\c{c}u
sans pr\'{e}c\'{e}dent des tendances d\'{e}mographiques des oiseaux et
des plantes, et contribuent \`{a} la production d'indicateurs de
biodiversit\'{e} bas\'{e}e sur la science ouverte et la recherche
reproductible.
\end{altabstract}

\shortrunauthors

\maketitle

\vspace*{-.6pc}

\twocolumngrid

\end{noXML}

\section{Introduction}\label{sec1}

{
Human-induced pressures are at the root of the ongoing environmental
crisis \citep{Ceballosetal2015}, which results in strong biodiversity
losses in many taxa, including birds
\citep{BirdLifeInternational2022,Ingeretal2015}, insects
\citep{Hallmannetal2017}, mammals \citep{CeballosEhrlich2002} and
plants \citep{Vellendetal2017,Jandtetal2022}, and also has consequences
for ecosystem functioning \citep{Biesmeijeretal2006,Martinetal2019}. 
To deal with this situation, policymakers try to provide guidelines to
take action against the current biodiversity crisis. For instance, the
recent Kunming-Montreal Global Biodiversity Framework sets four
strategic goals and 23 targets to mitigate biodiversity loss by 2050
\citep{CBD2022}. In particular, the first set of targets aims at
\textit{reducing threats to biodiversity}, with Target~4 specifically
highlighting the need to ``ensure urgent management actions to
halt~human-induced extinction of known threatened species and~for the
recovery and conservation of species, in particular threatened species,
[and] to significantly reduce extinction risk''. This requires the
identification of species whose populations are declining and careful
monitoring of the conservation status of threatened species.

To fulfill these goals, there is an urgent need to reinforce our
knowledge of the current state and dynamics of biodiversity, as well as
to develop reliable indicators that quantify and qualify trajectories
in biodiversity changes. These indicators must encapsulate complex
phenomena in a simple, understandable, and non-ambiguous way, be easy
to communicate, ecologically relevant, sensitive to changes, and
possess suitable mathematical properties, especially high precision and
absence of bias \citep{Gregoryetal2005,VanStrienetal2012}. In
view of these characteristics, estimated temporal trends in species
abundance or occurrence have been proposed as relevant indicators at
the population scale \citep{Pereiraetal2013}. However, many
{methodological} choices in the production of these trends remain
arbitrary and vary significantly across observatories, which can lead
to inconsistencies and hinder their comparability
\citep{Dornelasetal2013,Toszogyovaetal2024}.

The estimation of temporal trends relies on long-term monitoring of
species abundance or occurrence. Such data are difficult to collect,
especially at large spatial scales, because of the required sampling
effort and associated costs. One way to tackle this issue is to use
citizen-based monitoring programs, which rely on non-professional
volunteers to collect biodiversity data using standardized protocols.
This approach has multiple benefits, including lowering the costs
\citep{Levreletal2010,Theobaldetal2015}, expanding the spatial
coverage, while also developing public engagement regarding
conservation matters \citep{Deguinesetal2020,Pococketal2018}.

Although concerns have been raised regarding the reliability of
citizen-based data, with a focus on knowledge gaps between participants
and variations in sampling effort over time and space
\citep{Dickinsonetal2010}, there are now recommendations on the most
suitable monitoring frameworks for temporal trend estimation
\citep{Pescottetal2015} and on statistical tools that facilitate data
analysis while possibly accounting for those biases
\citep{PannekoekVanStrien2005}. This may explain the success of such
programs, measured by the number of scientific studies derived from
citizen science \citep{Cooperetal2014,Fraisletal2022} as well as the
development of citizen-based spatial and temporal indicators (e.g., UK
plant atlas or European Common Bird Indicator) that are used by local
and international institutions \citep{Chandleretal2017,Fraisletal2020}.
\looseness=-1

In France, Vigie-Nature is a citizen science program that offers
multiple schemes for volunteers to observe and monitor biodiversity,
depending on their knowledge and taxa of interest 
\citep[\url{http://www.vigienature.fr};][]{Julliard2017}.
The French \mbox{Breeding} Bird Survey (called STOC, the French acronym for
Suivi Temporel des Oiseaux Communs), which focuses on common bird
monitoring by skilled volunteer ornithologists, is one of the oldest
schemes (created in 1989). It is already used both nationally and
across Europe to inform on trends in bird populations, depending on
habitat specialization \citep{Fontaineetal2020}. Vigie-flore is a
more recent scheme that started in 2009, involving a large community of
skilled volunteer botanists to record the presence of vascular plants.
Although no population trends have yet been estimated from this
monitoring program, it has been used to assess the effects of climate
change on plant community composition \citep{Martinetal2019}.\looseness=-1
}

The production of annual population trends is expected by funding
institutions, NGOs involved in biodiversity conservation, and
participants. Furthermore, population abundances and their trends are
part of the Essential Biodiversity Variables (EBV) used to set IUCN
conservation statuses; they should hence be produced using open, robust
and reproducible statistical methodologies. Some tools already exist
for producing population trends based on count data, like the one in 
\citep{PannekoekVanStrien2005}, and are commonly used for taxa such as
birds (see the Pan-European Common Bird Monitoring Scheme (PECBMS) 
[\url{https://pecbms.info}]). However, they are not yet suitable for
analyzing occurrence data, such as those obtained from plant monitoring
in Vigie-flore. Beyond the choice of relevant statistical approaches,
producing population trends from citizen science data involves
fundamental steps of data processing that need to be standardized and
made available.\looseness=-1

Here, we present the analysis pipeline we developed to estimate
population trends, handling various data types (such as counts and
presence/absence), and taking into account the protocol specificities
of each monitoring scheme. We also present the resulting species
population trends for 148 common bird and 181 common plant species in
France, based on data collected through the STOC and Vigie-flore
monitoring schemes over the past 35 and 15 years, respectively. We
further introduce new criteria for trend characterization and
classification. Finally, we provide new graphical representations of
results to improve outreach, from understanding uncertainty to
non-linearity.

\vspace*{-4pt}

\section{Methods}

\vspace*{-3pt}

\subsection{Data}

\vspace*{-3pt}

\subsubsection{Bird abundance monitoring scheme}

\vspace*{-2pt}

Bird abundance data were produced by the STOC, a citizen science scheme
that has been monitoring common birds in France since 1989. This
program relies on a network of skilled volunteer ornithologists who are
assigned a $2\,{\times}\,2$~km square sampled from a systematic grid.
Each square contains 10~counting points that are representative of the
habitats found within the square. Each year, volunteers record all bird
individuals seen or heard during five minutes at each of these counting
points. Each year, there are two recording sessions during the breeding
period, separated by four to six weeks. The early session is held
before  May 8{th} to sample the early singing birds, and the late
session is held after May~8th, to sample the late migrants. To record
only birds that are actual breeders on the monitoring sites, the
migratory species that are still on their way to Northern Europe at
that time (yellow wagtail \textit{Motacilla flava}, whinchat
\textit{Saxicola rubetra}, meadow pipit \textit{Anthus pratensis},
northern wheatear \textit{Oenanthe oenanthe}, willow warbler
\textit{Phylloscopus trochilus} and fieldfare \textit{Turdus pilaris})
are excluded from the dataset of the first session. Recording sessions
must occur between one and four hours after sunrise and must be
repeated around the same dates every year.


The STOC protocol changed in 2001: before this date, squares were
chosen by volunteers. Since 2001, they have been randomly assigned
within a 10~km buffer around a location provided by volunteers. This
sampling design is more representative of habitats across France. Also,
the number of counting points per square ranged from five to 15 before
2001, but was fixed to 10 afterwards. Finally, the number of mandatory
visits per square and year increased from one to two in 2001. A full
description of the protocol is available on the Vigie-Nature website 
[\url{http://www.vigienature.fr}]. Due to this
change in protocol, we generally discarded the 1989--2000 data, except
for the estimation of an historical trend (see below).

We analyzed bird counts at the year and square levels. For each year,
species, and counting point, we retained the maximum number of
individuals recorded over the two sessions, and then summed these
values for all points within a square. This sum was used as a proxy for
the species abundance in a given square and year.

\begin{table*}%tab1
\caption{\label{tab1}Grouping of the CLC 44 land-use classes into five
broad classes\vspace*{-2pt}}
\begin{tabular}{ll}
\thead
\multicolumn{1}{c}{Broad classes} & \multicolumn{1}{c}{Original 44 classes} \\
\endthead
\morerows{1}{\textit{Forest cover}} & Level 3.1-Forests (\textit{3 classes}) \\
& Level 3.2-Shrub and/or herbaceous vegetation associations (\textit{4 classes})\vspace*{4pt} \\
\textit{Agricultural cover} & Level 2-Agricultural areas (\textit{11 classes}) \\
\textit{Urban cover} & Level 1-Artificial surfaces (\textit{11 classes}) \\
\morerows{1}{\textit{Water cover}} & Level 4-Wetland (\textit{5 classes}) \\
&  Level 5-Water bodies (\textit{5 classes})\vspace*{4pt} \\
\textit{Other covers} & Level 3.3-Open spaces with little or no vegetation (5 classes)
\botline
\end{tabular}
\vspace*{-2pt}
\end{table*}

As volunteers are, by definition, free to join and quit the monitoring
program at any time, squares were monitored over a variable number of
years. Over the 1989--2023 period of interest, 886 squares (24\%) were
monitored for a single year, and on average squares have been monitored
for 5.9 years. For the analysis, we only retained squares that were
visited for at least two years, as squares monitored for a single year
are not informative for population trends. This led to a total of 
21~010 visits of 2815 squares, of which 1134 visits of 184 squares
occurred before the 2001 change in protocol.

We analyzed the 148 most common species, defined as species observed in
at least five squares per year on average. As only presence 
\mbox{(and
abundance)} data are registered by volunteers, we reconstructed absence
data using the following procedure: a species observed at least one
year in a square was assigned an abundance of zero for all years in
which the species was not recorded, although the square was monitored.


\subsubsection{Plant sampling protocol}

Plant occurrence data were produced by Vigie-flore, a French citizen
science program that monitors wild flora in France since 2009. This
program relies on a network of skilled volunteer botanists who are
randomly assigned a $1\, {\times}\,1$~km square sampled from a
systematic grid. Each square contains eight systematically distributed
10~m$^{2}$ plots, which consist of 10 contiguous quadrats of 1~m$^{2}$.
Each year, volunteers record all vascular plant species that are
present in each quadrat. The proportion of quadrats in which a species
is observed (species frequency) is used as a proxy for its abundance in
each plot. The monitoring period is chosen depending on the square
location: squares located in the Mediterranean biogeographical region
are monitored between April~1st and May 31st; those located in
continental or Atlantic regions are monitored between June and July;
and squares located in areas with an elevation above 1000 meters
(alpine region) are monitored between July and August. A full
description of the protocol is available on the Vigie-Nature website 
[\url{http://www.vigienature.fr}].

For the same reasons as for the bird dataset, squares were monitored
for a variable number of years: 277 squares (41.6\%) were monitored for
a single year, and, on average, squares were monitored for four years.
For the analysis, we retained squares that were visited for more than
one year, meaning 2339 visits of 388 squares.



We analyzed the 181 most common species, defined as species observed in
at least ten squares per year on average. This threshold was set higher
than that for birds to compensate for the lower precision of
presence/absence data regarding count data. Absences were reconstructed
using the same methodology as for the bird dataset.

\begin{table*}%tab2
\caption{\label{tab2}Variables included as fixed effects in the models
for trend estimation\vspace*{2pt}}
\begin{tabular}{cclcc}
\thead
{Variables name} & {Type} & \multicolumn{1}{c}{Description} & {STOC} & {Vigie-flore} \\
\endthead
\textit{Year} & int & Observation year & X & X \\
\textit{Forest\_cover} & num & Percent forest cover & \morerows{3}{X}  &   \\
\textit{Urban\_cover} & num & Percent urban cover &  &  \\
\textit{Water\_cover} & num & Percent water cover &  &  \\
\textit{Other\_cover} & num & Percent other covers (excluding agricultural cover)\vspace*{4pt} &  &  \\
\textit{Longitude} & num & Longitude coordinates, treated as a 2{nd} order polynomial & \morerows{3}{X}  &   \\
\textit{Latitude} & num & Latitude coordinates, treated as a 2{nd} order polynomial &  &  \\
\textit{Longlat} & num & Interaction between longitude and latitude coordinates &  &  \\
\textit{Elevation} & int & Elevation\vspace*{4pt} &  &  \\
\textit{Date} & int & Date of observation, in a Julian date format &  & X \\
\textit{Visit} & char & \parbox[t]{9cm}{\raggedright Recording session (three categories: early session, late session, both sessions)}\vspace*{2pt} & X &
\botline
\end{tabular}
\vspace*{2pt}
\end{table*}

\subsubsection{Land cover data}

We characterized land cover on Vigie-flore plots and STOC squares using
the CORINE Land Cover (CLC) inventory \citep{Bossardetal2000}, which
relies on visual interpretation of high-resolution satellite imagery to
classify land cover into 44 classes. Four updates were made in 2000,
2006, 2012 and 2018. STOC squares and Vigie-flore plots were thus
associated with the closest previous CLC update: CLC 2018 for squares
and plots visited since 2018, CLC 2012 for squares and plots visited
between 2012 and 2018 (excluded), and so on.


We categorized the 44~land cover classes into five broad classes:
forest, agricultural, urban, water and other land cover (see 
Table~\ref{tab1}). For the bird dataset, land cover was characterized
through the percent cover of each of the five broad classes found in a
1~km buffer around the center of each STOC square. The percentage of
agricultural cover was removed from the analysis to avoid
multicollinearity in models, as all percentages sum to 100. For the
plant dataset, land cover was characterized through a five-class
variable corresponding to the major land cover (i.e., land cover with
the largest area) found within a \mbox{5-meter} buffer around the center of
each \mbox{Vigie-flore} plot. This buffer size was chosen so that land cover
could be described locally. The choice of a categorical variable is
justified by the negligible number of plots with multiple land cover
classes within the buffer.

\vspace*{-3pt}

\subsection{Estimation of temporal trends in bird abundance or in plant
probability of occurrence} 
\vspace*{-3pt}

\subsubsection{Population trends estimation}

For each species in the plant and bird datasets, we estimated a linear
population trend over the entire monitoring period (2001--2023 for
birds and 2009--2023 for plants), as well as a linear population trend
over the last ten years (2014--2023). To do so, we fitted a generalized
linear mixed model for each species, using the R package glmmTMB
\citep{Brooksetal2017}, following previous work on similar datasets
\citep{SauerLink2011}. Covariates were introduced to correct for
possible biases in trend estimation, following a variable selection
procedure based on Bayesian Information Criteria (Supplementary
material ``Variable Selection Procedure''). For the plant dataset, the
only covariate retained was the observation date, in a Julian date
format. For the bird dataset, the recording session was included as a
three-level variable: early session, late session or both sessions.
Land cover (described above) was also retained. For collinearity
matters, we excluded agricultural cover. Finally, elevation was also
added as a fixed effect, as well as \mbox{longitude} and latitude, which were
both treated as 2{nd} order polynomial effects, and their interaction
(see Table~\ref{tab2}).


Random effects were also introduced in those models, to take dependence
among observations into account. This dependence is mostly due to
repeated observations in the same squares and/or by the same
volunteers. For birds, we included a random intercept for squares
nested within administrative d\'{e}partements. For plants, we included
a random intercept for plots nested within squares, as well as a random
intercept for volunteers. For both datasets, we also added a random
effect of plots on the slope of the year effect.

For bird abundance, we used a negative binomial distribution with a
quadratic parametrization as well as a log link function, designed to
handle overdispersion in count data. For plant occurrence, we used a
beta-binomial distribution with a clog-log link function, designed to
handle overdispersion in presence/absence data.

\subsubsection{Annual variations}

To estimate annual variations in bird abundance and plant occurrence,
we fitted a generalized linear mixed model with similar specifications
as above, but with two major changes. First, the variable \textit{year}
was treated as a categorical variable rather than a continuous
variable. This allows for the estimation of the difference in mean
abundance between each year and a reference level. This reference level
is usually set to the abundance of the first year, but due to the low
number of observations during the first year of both monitoring
schemes, we set the reference level to the mean abundance across all
years, thus avoiding convergence issues. Second, as \textit{year} is
treated as a categorical variable, the models did not include random
slopes for the year effect across\break squares/plots.

For visualization and communication purposes, we smoothed annual
variations by fitting a generalized additive mixed model (GAMM), using
the R package mgcv \citep{Wood2017}. Following \citep{Fewsteretal2000},
the number of knots ${k}$ was set to 30\% of the time series, thus
authorizing a changepoint every three years on average. As smoothing
was used for illustrative purposes, we simplified the model
specifications by removing random slopes. Also, due to restrictions on
implemented distribution laws, we used a quasi-binomial distribution
law to model plant occurrences, which also accounts for overdispersion
in presence/absence data.

\subsubsection{Historical trends}

For each species in the bird dataset, we also estimated a linear
population trend over the monitoring period that covers both protocols
(1989--2023), thus adding historical data collected with the 1989--2001
protocol. Due to changes in protocol and a smaller number of
observations before 2001, we simplified the model calibrated on the
2001--2023 period and used different covariates from those described
above. First, proportions of water cover and other covers were removed
from the covariates. Because the number of monitoring points was
variable before 2001, we included it as a covariate in the model. We
used the log number of monitoring points to keep proportionality
between the number of points and abundance despite the log link
function. To account for differences in protocols before and after
2001, we added the corresponding two-class variable as a fixed effect.
Finally, recording sessions in the early protocol were not repeated
twice, and no information on the timing of the single session was
available (early or late session in the description of the bird
monitoring protocol). To incorporate this information into the model
nonetheless, we randomly assigned a timing for the recording session of
each site visited before 2001 (early or late session, both with
probability 0.5). This timing of sessions was fixed across the years.
\mbox{Finally}, random effects were the same as\break above.

\subsubsection{Uncertainty curves}

\vspace*{2pt}

To better communicate on uncertainty, we relied on
\citep{Pescottetal2022} who suggest the use of ensemble lines to
represent a range of possible trajectories that are compatible with
medium-term and short-term trends. For each species $s$, an
ensemble line was built using the following procedure:
\begin{itemize}
\item  Relative abundance means $\beta _{s}= ({\beta }_{s}^{1},\break
{\beta }_{s}^{2},\ldots,{\beta }_{s}^{n})$ and standard
errors $se_{s}=({se}_{s}^{1},{se}_{s}^{2},\ldots ,{se}_{s}^{n})$
associated with each year $i=\{1,\ldots ,n\}$ were
extracted from annual variation models;
\item We drew 100 simulated sets of yearly relative abundance
estimates $X_{j,s}=({x}_{j,s}^{1},{x}_{j,s}^{2},\ldots
,{x}_{j,s}^{n})$, $j=\{1,\ldots ,100\}$ from a normal
distribution with the following parametrization: $X_{j,s}\sim N(\beta
_{s},se_{s})$;
\item  For each simulated set of estimates, a linear regression was fitted;
\item  Intercepts and slopes were then extracted from each of these
regressions to draw a line ensemble that shows 100 possible
trajectories.\vspace*{2pt}
\end{itemize}

\subsubsection{Trends classification}

\vspace*{2pt}

For each species $s$, the estimated temporal trends in
abundance/occurrence ${\beta }_{s}^{\mathrm{year}}$ were converted to growth
rates, as follows: $GR_{s}=\exp ({\beta}_{s}^{\mathrm{year}})$. The standard
errors associated with the trend estimates, ${se}_{s}^{\mathrm{year}}$, were
used to infer the lower (${GR}_{s}^{\mathrm{inf}}$) and upper bounds
$({GR}_{s}^{\mathrm{sup}})$ of the growth rate confidence interval:
\vspace*{2pt}
{\begin{eqnarray*}
{GR}_{s}^{\mathrm{inf}} &=&\exp({\beta }_{s}^{\mathrm{year}}-1.96\times {se}_{s}^{\mathrm{year}});\\
{GR}_{s}^{\mathrm{sup}} &=&\exp ({\beta }_{s}^{\mathrm{year}}+1.96\times {se}_{s}^{\mathrm{year}}).
\end{eqnarray*}\vspace*{2pt}}\unskip
We relied on the IUCN Red List framework to classify trends into
different categories. We specifically used the A2 criterion that states
a species should be treated as vulnerable in case of ``an observed,
estimated, inferred or suspected population size reduction of
${\geq}$30\% over the last ten years or three generations, whichever
is the longest, where the reduction or its causes may not have ceased
OR may not be understood OR may not be reversible [{\ldots}]''
\citep{IUCN2012}.

Therefore, we used as lower (resp.\ upper) limits of our classification
scheme, growth rates that matched a decline (resp.\ increase) of more
than 30\% in ten years, as lower (resp.\ upper) limits of our
classification scheme. That is: $GR_{\mathrm{lower}\_\lim}=0.961$ and
$GR_{\mathrm{upper}\_\lim}=1.030$. This yielded the exclusive
categories that can be found in Figure~\ref{fig1}.

\begin{figure}[b!]
\includegraphics{fig01}
\vspace*{-2pt}
\caption{\label{fig1}Linear-trend classification depending on estimated growth
rates. (a) Strong increase: the lower bound of the confidence interval
is larger than the upper limit
($GR_{s}^{\mathrm{inf}}>GR_{{\mathrm{upper}\mathop{\_}\lim}}$); (b)
Moderate to strong increase: the lower bound of the confidence interval
is larger than 1 (significant increase) and the growth rate is larger
than the upper limit
($1<GR_{s}^{\mathrm{inf}}<GR_{\mathrm{upper}\mathop{\_}{\lim}}$ and 
$GR_{s}>GR_{\mathrm{upper}\mathop{\_}{\lim}}$); 
(c) Moderate increase: the
lower limit of the confidence interval is larger than 1 (significant
increase) and the upper bound of the confidence interval is smaller
than the upper limit ($1<GR_{s}^{\mathrm{inf}}$ and 
$GR_{s}^{\mathrm{sup}}< GR_{\mathrm{upper}\mathop{\_}{\lim}}$); (d)
Stable: the confidence interval encloses 1 (no significant trend) and
is included within the interval  [$GR_{\mathrm{lower}\mathop{\_}\mathrm{lim}}$; 
$GR_{\mathrm{upper}\mathop{\_}{\lim}}$]; (e) Moderate decline: the upper
bound of the confidence interval is smaller than 1 (significant
decline) and the lower bound of the confidence interval is larger than
the lower limit; (f) Moderate to strong decline: the upper bound of the
confidence interval is smaller than 1 (significant decline) and the
growth rate is smaller than the lower limit
($GR_{\mathrm{lower}\mathop{\_}\mathrm{lim}}>GR_s$ and $GR_s^{\mathrm{sup}}<1$);
(g) Strong decline: the upper bound of the confidence interval is
smaller than the lower limit
($GR_{s}^{\mathrm{sup}}<GR_{\mathrm{lower}\mathop{\_}\mathrm{lim}}$); (h)
Uncertain: the confidence interval encloses 1 (no significant trend)
and encloses either $GR_{\mathrm{lower}\mathop{\_}\mathrm{lim}}$ or 
$GR_{\mathrm{upper}\mathop{\_}{\lim}}$.\looseness=1}
\vspace*{-2pt}
\end{figure}

\subsubsection{Population trends within habitat specialization groups}

\vspace*{-2pt}

To better understand variations in species trends, we examined
population trends within habitat specialization categories,
{sensu} \citep{Jiguetetal2012}. We distinguished trends in
generalist, rural specialist, farmland specialist, urban specialist,
and other common species, hereafter referred to as ``others'' 
(Tables~\ref{tab3} and~\ref{tab4}; Supplementary material ``Habitat
Specialization'').

\begin{table*}%tab3
\caption{\label{tab3}Partitioning of bird long-term trends (2001--2023) within the different habitat specialization categories}
\begin{tabular}{cccccc}
\thead
& {Generalists} & \parbox[t]{2cm}{\centering Farmland specialists} & 
\parbox[t]{2cm}{\centering Forest specialists} & 
\parbox[t]{2cm}{\centering Urban specialists} & 
\parbox[t]{1.5cm}{\centering Other species}\vspace*{2pt} \\
\endthead
{Strong increase} & 7.1\%  (1) & 0\% & 4.2\%  (1) & 7.7\%  (1) & 12.7\%  (9) \\
{Moderate to strong increase} & 0\% & 0\% & 4.2\%  (1) & 0\% & 7\%  (5) \\
{Moderate increase} & 21.4\%  (3) & 16.7\%  (4) & 25\%  (6) & 30.8\%  (4) &  11.3\%  (8) \\
{Stable} & 50\%  (7) & 37.5\%  (9) & 16.7\%  (4) & 7.7\%  (1) & 23.9\%  (17) \\
{Moderate decline} & 21.4\%  (3) & 20.8\%  (5) & 41.7\%  (10) & 30.8\%  (4) & 11.3\%  (8) \\
{Moderate to strong decline} & 0\% & 8.3\%  (2) & 4.2\%  (1) & 7.7\%  (1) & 7\%  (5) \\
{Strong decline} & 0\% & 12.5\%  (3) & 4.2\%  (1) & 15.4\%  (2) & 1.4\%  (1) \\
{Uncertain} & 0\% & 4.2\%  (1) & 0\% & 0\% & 25.4\%  (18) 
\botline
\end{tabular}
\end{table*}

\begin{table*}%tab4
\caption{\label{tab4}Partitioning of the short-term trends (2014--2023)
within the different habitat specialization categories}
\vspace*{-5pt}
\begin{tabular}{cccccc}
\thead
& {Generalists} & \parbox[t]{2cm}{\centering Farmland specialists} & 
\parbox[t]{2cm}{\centering Forest specialists} & 
\parbox[t]{2cm}{\centering Urban specialists} &
\parbox[t]{1.5cm}{\centering Other species}\vspace*{2pt} \\
\endthead
{Strong increase} & {7.1\%  (1)} & {12.5\%  (3)} & {8.7\%  (2)} & {15.4\%  (2)} & {19.4\%  (14)} \\
{Moderate to strong increase} & 0\% & {4.2\%  (1)} & {8.7\%  (2)} & {15.4\%  (2)} & {11.1\%  (8)} \\
{Moderate increase} & {35.7\%  (5)} & {12.5\%  (3)} & {20.9\%  (5)} & {28.6\%  (4)} & 0\% \\
{Stable} & {28.6\%  (4)} & {16.7\%  (4)} & {17.4\%  (4)} & {7.7\%  (1)} & {6.9\%  (5)} \\
{Moderate decline} & {28.6\%  (4)} & {4.2\%  (1)} & {20.9\%  (5)} & {28.6\%  (4)} & {6.9\%  (5)} \\
{Moderate to strong decline} & 0\% & {20.8\%  (5)} & {8.7\%  (2)} & 0\% & {9.7\%  (7)} \\
{Strong decline} & 0\% & {4.2\%  (1)} & {8.7\%  (2)} & {7.7\%  (1)} & {1.4\%  (1)} \\
{Uncertain} & 0\% & {25\%  (6)} & {8.7\%  (2)} & {7.7\%  (1)} & {44.4\%  (32)} 
\botline
\end{tabular}
\vspace*{-7pt}
\end{table*}

\begin{figure*}
\includegraphics{fig02}
\caption{\label{fig2}Bird population estimated growth rates and trend
classification for the 2001--2023 (A and B) and the 2014--2023 (C and
D) periods. In panels A and C, error bars around bird population
estimated growth rates represent 95\% confidence intervals. Histograms
on panels B and D indicate the number of bird species per class of
population trend.}
\vspace*{4pt}
\end{figure*}

\vspace*{-2pt}

\subsubsection{Code and data availability}

The data and the R code used for the presented analysis is available at
\url{https://zenodo.org/records/14957111}. As the code is under continuous
development, the most up-to-date version of the code is available at
\url{https://outils-patrinat.mnhn.fr/gitlab/vigie-nature/tendances/indicatorroutine}.

\vspace*{-3pt}
\section{Results}
\vspace*{-3pt}

\subsection{Bird species population trends}

\vspace*{-2pt}

The estimation of bird population trends over the 2001--2023 period
converged for 146 out of 148~species. The estimated trends in abundance
range from ${-}$0.093 (95\% confidence interval: CI~${=}$~[${-}$0.109; ${-}$0.076])
for the European tree sparrow \textit{Passer montanus,} corresponding
to an annual growth rate of 0.911 (CI ${=}$ [0.897; 0.926]), to 0.206
(CI~${=}$~[0.169; 0.242]) for the rose-ringed parakeet \textit{Psittacula
krameria} i.e., an annual growth rate of 1.229 (CI~${=}$~[1.184; 1.274]).
These extremes represent respectively a mean decrease in abundance of
87.1\% and a mean increase in abundance of 9188\% over the 23-year
monitoring period.


Regarding population trend classification, more than one third of
trends over the 23-year period are classified either as ``stable'' (38
species) or ``uncertain'' (19 species), see Figure~\ref{fig2}.
Declining and increasing trends are evenly distributed, with 43 species
showing increasing population trends, while 46 species population
trends are classified as \mbox{declining}. Regarding the intensity of increase
(resp.\ decline), the majority of species (25 out of 43) was associated
with population trends classified as ``moderately \mbox{increasing}'' (resp.\ 
``moderately decreasing'', 30~out of 46). All species estimated trends
and \mbox{classifications} over the 2001--2023 period are available in
\mbox{Appendix~1}.

\begin{figure*}
\includegraphics{fig03}
\vspace*{-5pt}
\caption{\label{fig3}Alluvial plot showing the relationships between
bird population trends over the 2001--2023 and the 2014--2023 periods.
Each ribbon corresponds to a bird species. Five species for which a
model did not converge for one of the periods and 18 species with
uncertain trends for both periods (i.e., transitions from ``Uncertain''
to ``Uncertain'') were not represented to improve readability.}
\vspace*{-3pt}
\end{figure*}


Over the shorter 2014--2023 period, model convergence was achieved for
145 out of 148 species. Trend estimates over the last ten years range
from ${-}$0.169 (CI~${=}$~[${-}$0.215; ${-}$0.124]) for the Eurasian tree
sparrow \textit{Passer montanus}, corresponding to an annual growth
rate of 0.844 (CI~${=}$~[0.807; 0.883]), to 0.291
(CI~${=}$~[0.184; 0.398]) for the western cattle egret
\textit{Bubulcus ibis}, i.e.\ an annual growth rate of 1.338
(CI~${=}$~[1.20Z; 1.489]). These extremes represented respectively a
decrease of 78.2\% and an increase of 1274\% over the ten-year
monitoring period. Standard errors for these trend estimates are on
average 0.0191, which is more than two times larger than for trend
estimates over the 2001--2023 period (0.0079).

Patterns of trend classification are similar for the last ten years and
for the 23-year period (Figure~\ref{fig2}). A little more than a third
of trends over the last ten years are classified as either ``stable''
(18 species) or ``uncertain'' (41 species), though we note an increase
in the number of uncertain trends that is consistent with both the
decrease in the available number of observations and the resulting
increase in standard errors. As for the 23-year period, declining and
increasing trends are evenly distributed (50 increasing species
population trends against 38 declining). However, moderate trends tend
to be less frequent, with 17 ``moderately increasing'' species
population trends (resp.\ 19 ``moderately decreasing'' trends) compared
to the 23-year period. Also, in the last ten years, there has been a
reinforcement of species with a strong increase in abundance (from 12
to 22 species) and no such change regarding strong declines in
abundance (from seven to five species). All species estimated trends
and classifications over the 2014--2023 period are available in
Appendix~2.

\begin{figure*}
\includegraphics{fig04}
\vspace*{-6pt}
\caption{\label{fig4}Examples of temporal trends in abundance for four
bird species. The orange line corresponds to the temporal trend
smoothed over the 2001--2023 period, with the orange ribbon
representing uncertainty. Orange dots correspond to ratios of abundance
between each year and the reference year. Blue and red lines correspond
to uncertainty associated with the population trend estimated over the
2001--2023 period (resp.\ 2014--2023 period). Blue and red numbers on
the left of the plot indicate the estimated annual growth rate for the
2001--2023 period (resp.\ 2014--2023 period). Stars on the right of the
plot specify the significance of these estimates.}
\vspace*{-3pt}
\end{figure*}


Estimating linear trends over a long period (2001--2023), however,
masks more complex temporal patterns in abundance, which can be
uncovered by \mbox{comparing} long (2001--2023) vs.\ shorter (2014--2023) term
trends, or more finely by examining the non-linear patterns illustrated
by the outputs of the GAMM (Appendix~6). Looking at the changes in
trend \mbox{classification} between the 2001--2023 and the 2014--2023 periods
for the 125 species for which models converged for both periods and for
which trends were not ``uncertain'' in both periods (Figure~\ref{fig3}),
we found that species frequently changed trend classes across the two
periods, sometimes with a change in the direction of the trend, or with
a jump of one or more classes (e.g, from ``moderate increase'' to
``strong increase''). These changes thus covered a variety of
situations, from accelerated declines (e.g.\ the Wood warbler
\textit{Phylloscopus sibilatrix}) to recent recoveries, e.g.\ the
Goldfinch \textit{Carduelis carduelis}, or Cetti's warbler
\textit{Cettia cetti}, for which the temporal trends in abundance
produced by the GAMM are strongly non-linear (Figure~\ref{fig4}). For
Cetti's warbler, examining the 2001--2023 linear trend thus says
nothing of the initial decline between 2001 and 2012. \mbox{Finally}, some
species show no change in trend classes between the two \mbox{periods}, as is
the case with the long-term stability of the Carrion crow
\textit{Corvus corone} (Figure~\ref{fig4}).


An absence of classification change in estimated population trends is
not equivalent to an absence of absolute change in estimated population
trends. For example, though the Eurasian golden oriole \textit{Oriolus
oriolus} population trend is classified as ``moderately increasing''
for both periods, the estimated annual growth rate increased from 1.005
to 1.029 (see Appendix~1 and 2), suggesting a recent stronger increase
for this species. However, studying all changes in estimated trends
between 2001--2023 and 2014--2023, we found that differences in
estimates are centered around 0, with a slightly positive mean
\mbox{difference} of $8.68~{\times}~10^{-3}$ (CI~${=}$~[$3.33\times10^{-3}$; 
$1.40\times 10^{-2}$]) corresponding to a mean ratio of 1.0087
(CI~${=}$~[1.0033; 1.0141]).

Looking at population trends within groups of species habitat
specialization, we notice that, for the 2001--2023 period, most habitat
generalist species show ``stable'' (50\%) or ``moderately increasing''
(21.4\%) population trends, with no strong declines recorded. In
contrast, forest specialists display a high proportion of populations
with declining trends (50\%), relatively few with stable trends
(16.7\%) but still a substantial proportion whose trends are increasing
(33.3\%). Urban specialists show a similar bimodal pattern, with 54\%
of species displaying \unskip\break\mbox{declining} trends, while 38.5\% display increasing
trends. The majority of farmland specialist species populations are
declining (41.6\%) or stable (37.5\%) with relatively few increasing
population trends (16.7\% moderately increasing, and no strong
\mbox{increase)}.


These patterns slightly differ over the shorter 2014--2023 period
(Table~\ref{tab4}), with more numerous strongly increasing population trends
among \mbox{farmland} specialist species, and a larger number of uncertain
trends, particularly among ``other'' species.



When adding historical data collected from 1989 to 2001 with a
different protocol (see methods) to estimate population trends over the
period covering both protocols (1989--2023), model convergence was
achieved for 107 species only (Appendix~3). The estimated trends in
abundance are very similar to the ones obtained for the 2001--2023
period, with a mean difference of ${-}2.02\times 10^{-4}$
(CI~${=}$~[${-}7.94\times 10^{-4}$; $3.90\times10^{-4}$]) and a maximum
absolute difference of $1.51\times 10^{-2}$. This is almost ten times
less than the maximum difference observed between the 2001--2023 and
the 2014--2023 periods ($1.37\times 10^{-1}$), suggesting that
observations made before 2001 contribute only marginally to the
1989--2023 estimated trends, mainly due to their scarcity compared to
the amount of data collected since 2001.

\begin{figure*}
\includegraphics{fig05}
\vspace*{-3pt}
\caption{\label{fig5}Plant population trend estimates and
classification for the 2009--2023 (A and B) and the 2014--2023 (C and
D) periods. In panels A and C, error bars associated with plant
population trend estimates represent 95\% confidence intervals.
Histograms on panels B and D indicate the number of plant species per
class of population trend.}
\vspace*{-3pt}
\end{figure*}


\subsubsection{Plant species population trends}
\vspace*{-5pt}

The estimation of plant population trends over the 2009--2023 period
reached convergence for all 181 plant species. These trend estimates
ranged from ${-}$0.106 (CI~${=}$~[${-}$0.197; ${-}$0.015]) for the common box
\textit{Buxus sempervirens} corresponding to a growth rate of 0.899
(CI~${=}$~[0.821; 0.984]), to 0.100 (CI~${=}$~[0.053; 0.147]) for the
dandelion \textit{Taraxacum officinale} corresponding to a growth rate
of 1.105 (CI~${=}$~[1.055; 1.158]). These extremes represented
respectively a decrease of 77.4\% and an increase of 307.7\% over the
15-year monitoring period. Standard errors for these trend estimates
were on average 0.0256, which is about three times larger than for bird
species.

Most plant species trends over the 15-year period were classified as
``uncertain'' (126 species), i.e., the estimation was not precise
enough to conclude for 70\% of species. For the remaining species, we
observed fewer increasing than declining population trends
(respectively 15 and 32 species) as well as an over-representation of
trends categorized as ``moderately to strongly'' increasing (9 species)
or declining (23 species). All species estimated trends and
{classifications} over the 2009--2023 period are available in Appendix~4.


Turning to the estimation of plant population trends over the last ten
years, model convergence was achieved for all 178 species out of 181
(Figure~\ref{fig5}). These trend estimates over the last ten years
ranged from ${-}$0.209 (CI~${=}$~[${-}$0.300; ${-}$0.117]) for the tor-grass
\textit{Brachypodium pinnatum} corresponding to an annual growth rate
of 0.811 (CI~${=}$~[0.740; 0.889]), to 0.235
(CI~${=}$~[0.139; 0.332]) for the dandelion \textit{Taraxacum
officinale} corresponding to an annual growth rate of 1.265
(CI~${=}$~[1.149; 1.394]). These extremes represented respectively a
decrease of 84.8\% and an increase of 731.7\% over the last ten-year
monitoring period. Standard errors for these trend estimates were on
average 0.0405, which is around 1.5 times larger than for trend
estimates over the 2009--2023 period. All species estimated trends and
classifications over the 2014--2023 period are available in \mbox{Appendix~5}.



The pattern obtained in terms of trend classification for the last ten
years did not differ much from the one obtained for the 15-year period,
except that most species population trends categorized as ``stable''
over the 2009--2023 period switched to ``uncertain'' over the last
ten-year period and that some species switched from an ``uncertain''
trend to a ``moderate to strong increase'' or ``moderate to strong
decline'' trend (Figure~\ref{fig5}). Yet, when comparing trend
classification across the two time periods, the few changes in trend
classes were diverse, as was the case for birds (Figure~\ref{fig6}),
including one accelerated increase (the English ryegrass \textit{Lolium
perenne}, Figure~\ref{fig7}), long-term stability (e.g.\ blackberry,
\textit{Rubus fruticosus}, Figure~\ref{fig7}) or transition to
``uncertain'' trends (e.g.\ The Traveller's Joy \textit{Clematis
vitalba}, Figure~\ref{fig7}). As for birds, the smoothed non-linear
trends uncover finer temporal patterns (Appendix~7), such as the recent
decline in a species population that appeared to be formerly stable
over long- and short-term trends (\textit{Buxus sempervirens},
Figure~\ref{fig7}).

\begin{figure*}
\includegraphics{fig06}
\vspace*{-4pt}
\caption{\label{fig6}Alluvial plot showing the relationships between
plant population trends over the 2009--2023 and the 2014--2023 periods.
Each ribbon corresponds to a plant species. To improve readability, the
106 species with uncertain trends for both periods (i.e., transitions
from ``uncertain'' to ``uncertain'') were not represented.}
\vspace*{-3pt}
\end{figure*}

\begin{figure*}
\includegraphics{fig07}
\caption{\label{fig7}Examples of temporal trends in abundance for four
plant species. The orange line corresponds to the temporal trend
smoothed over the 2009--2023 period, with the orange ribbon
representing uncertainty. Orange dots correspond to ratios of abundance
between each year and the reference year. Blue and red lines correspond
to uncertainty associated with the population trend estimated over the
2009--2023 period (resp.\ 2014--2023 period). Blue and red numbers on
the left of the plot indicate the estimated annual growth rate for the
2009--2023 period (resp.\ 2014--2023 period). Stars on the right of the
plot specify the significance of these estimates.}
\end{figure*}

\section{Discussion}

In this paper, we present an analysis pipeline for estimating the
population trends of multiple species across different taxa and time
periods. This pipeline complements existing tools, in particular TRIM~3
\citep{PannekoekVanStrien2005}, by being more versatile for various
data types, including occurrence data. We demonstrate this in our
analysis of plant species. While we did not implement all TRIM~3
features, such as observation weighting, or missing data imputation, we
leveraged mixed models and the glmmTMB package \citep{Brooksetal2017},
offering users flexibility in model parametrization. We also focused on
result visualization and interpretation, proposing new graphical
representations of population trends that include a trend over the
entire period of the data, a trend over the last ten years, and a
smoothed trend to capture non-linear or even non-monotonous patterns.
These figures are automatically generated for each species if models
converge. Drawing from the IUCN Red List framework \citep{IUCN2012} and
the PECBMS classification, we developed a seven-class categorization
that is better suited for temporal trend analysis. Our classification
is more sensitive to uncertainty and reduces the risk of \mbox{biologically}
irrelevant abrupt category changes. Finally, for transparency, we saved
all raw models and main results in tables, so that users can access
them beyond visual representations.
\vspace*{-3pt}

\subsection{Overview of population trends in birds and plants}

The analysis pipeline that we propose uses data from citizen science
monitoring schemes targeting bird and plant species, and provides an
unprecedented overview of the current dynamics of common biodiversity
in France. While annual bird population trends were already produced
from the same dataset using TRIM~3  \citep{Fontaineetal2020}, 
this is the first time that plant population trends have been made
available. Looking at the population trends over the entire monitoring
periods (2001--2023 for birds and 2009--2023 for plants), we found
almost as many bird species with declining populations as those with
{increasing} ones, and slightly more plant species with declining
populations than with \mbox{increasing} ones. For~bird species, these results
may appear to contrast with recent studies highlighting bird abundance
decline in Europe or France \citep{Lehikoinenetal2019,Rigaletal2023}.
Such studies usually tend to focus on smaller sets of bird species
consisting of habitat specialists (e.g.\ farmland, woodland, or mountain
species), for which it is worth noting that we also found a majority of
declining trends (Tables~\ref{tab3} and~\ref{tab4}). 
In contrast, our analysis covers all
common bird species in France, including habitat generalists and exotic
species with population sizes that are known to be mostly stable or
increasing \citep{LeVioletal2012}, similarly to what we found. Such a
pattern of habitat generalist species showing increasing populations
while habitat specialist species show declining ones is consistent with
biotic homogenization \citep{Devictoretal2008}. However, our results
further show that habitat specialization alone does not explain the
variations in bird population trends. Indeed, although we found that 11
out of 14 habitat generalist bird species had stable or increasing
populations, the remaining three species showed declining abundances
(Dunnock \textit{Prunella modularis}, Green woodpecker \textit{Picus
viridis} and Common cuckoo \textit{Cuculus canorus}; see Appendix~1
and Supplementary material ``Habitat Specialization''). Similarly,
among habitat specialist species, 29 show decreasing trends (e.g.\ 
Meadow pipit \textit{Anthus pratensis} or Tree sparrow
\textit{Passer montanus}), but 17 population trends are classified as
increasing (e.g.\ Jackdaw \textit{Corvus monedula} or Hawfinch
\textit{Coccothraustes coccothraustes}) and 14 as stable (e.g.\ Hoopoe
\textit{Upupa epops} or Kestrel \textit{Falco tinnunculus}). These
results therefore suggest that we need to go \mbox{beyond} habitat
specialization to fully understand bird population trends. We should
consider other potential response traits, such as trophic
specialization \citep{Bowleretal2019}.

An important caveat when interpreting temporal trends is that they
might be sensitive to changes in species detectability that are
associated with their life cycles. Thus, a change in phenology, which
has already been highlighted for birds \citep{Romanoetal2023} 
and for plants \citep{Piaoetal2019}, could lead to the detection of a
temporal trend in detectability, instead of an actual temporal change
in abundance or occurrence. To be precise, because the STOC protocol
encourages recording sessions to occur around the same dates every
year, a slight phenological advance or delay could lead to the
detection of an increasing (resp.\ decreasing) trend in abundance. Yet,
the potential impact of phenological changes on the resulting estimated
trends could be tested by incorporating the exact recording session
dates into the analysis, so as to model phenology in addition to trends
in abundance. Currently, this is done for plants, although phenology
might be better estimated using at least a quadratic effect on the
observation date. Others have done so, such as 
\citep{Schmuckietal2016} for butterflies or \citep{Moussusetal2009} for
birds.\looseness=-1

Turning to plants, the scarcity of studies on this topic in France
makes it difficult to compare our \mbox{results} to previous work (but see
\citep{Friedetal2009} for arable weeds). Nonetheless, our results are
consistent with patterns observed in other countries, such as more
numerous losses than gains in frequency for plants in Germany
\citep{Jandtetal2022}. Overall, we observed that many bird and plant
species exhibited either declining or increasing population trends.
This indicates that the composition of common bird and plant assemblages
is strongly changing, a result confirming previous studies, such as
\citep{Rigaletal2022} for birds and \citep{BuhlerRoth2011} for
plants. While several studies have documented such changes in French
bird communities \citep[e.g.][]{Devictoretal2008,Loreletal2021}, the
consequences for the composition of plant communities in France remain
to be\break examined.

\subsection{Non-linearity in population dynamics}

Our analysis pipeline, by estimating linear population trends over
multiple time periods, showed that trends are highly dependent on
temporal depths. Indeed both the direction and speed of species' growth
rates vary depending on the considered time-period. This is consistent
with recent studies that point out the non-linearity of temporal trends
in population abundance or occurrence, as well as the impact of the
chosen time period on these estimated trends \citep{Duchenneetal2022}.
Such non-linearities in population trends call for detailed and
species-specific interpretations. For instance, the population trend of
Cetti's warbler \textit{Cettia cetti} goes from ``moderate increase''
in 2001--2023 to ``strong increase'' during the 2014--2023 period. This
sedentary and insectivorous species is highly vulnerable to harsh
winters. It has experienced important population collapses in the past
after the 1984--1985 and 1986--1987 cold winters in France
\citep{IssaMuller2015}. It is likely that this species has benefited
from global warming and mild winters in recent years, such that its
range is shifting northward in Western Europe \citep{Clement2020}.
However, the explanation for the differences in trends observed over
longer vs.\ shorter time periods is not always straightforward: the
changing trends of the Wood warbler \textit{Phylloscopus sibilatrix}
may be attributable to a combination of factors, including changing
forest management, locally or on their wintering grounds, prey
availability, and climate change; however, the exact contribution of
each potential driver remains uncertain and needs to be investigated.
These examples all illustrate the interest of highlighting trends in
different timeframes to get a finer insight into bird population
dynamics.

We are aware that our current methodology, which consists of splitting
the timeline to cover either the whole monitoring period or the last
ten years, {cannot} perfectly encompass the dynamics of species
\mbox{populations.} As a first step, we chose to illustrate this non-linearity
using generalized additive models to calculate smoothed trends over the
entire period. One major interest of this trend representation is to
reveal complex population dynamics that cannot be properly described
using linear trends, even when considering different time periods. For
instance, the European hoopoe \textit{Upupa epops} is considered as
``stable'' with our categorization during both time periods. However,
its annual variations (Appendix~6, pp.~2) show remarkable cyclical
dynamics, with a period of seven years. This pattern could be related
to the population dynamics of one of its important prey 
\citep{BarbaroBattisti2011}, the pine processionary moth
\textit{Thaumetopoea pityocampa}, whose population outbreaks every
seven to nine years in France \citep{Lietal2015}. However, smoothing
trends were only implemented as a visual tool in the pipeline, and
further development is needed to properly characterize and quantify
this non-linearity . \citep{Rigaletal2020} for example, suggest the use
of a 2{nd} order polynomial regression, which allows for non-linearity
of the time effect, while still being understandable and biologically
relevant.

\subsection{Challenges and limitations in trend estimation}

With this non-linearity in mind, and given that pressures on
biodiversity have started long before the starting dates of
biodiversity monitoring schemes \citep{Mihoubetal2017}, one of our
objectives was to challenge our current 2001 baseline and study what
would happen if we went back further in time regarding bird population
trends. Our results, however, were impacted by the low number of
observations available before 2001, which caused trends to be mostly
driven by observations made after 2001. This finding was illustrated by
the small number of differences observed between estimated trends over
the 1989--2023 period and the 2001--2023 period. Such limitations
related to data availability are also highlighted by the higher
uncertainty associated with trends over the last ten years for both
bird and plant species. Similarly, the impact of sampling effort was
illustrated by the overall wider confidence intervals observed for
plant population trends compared to those observed for birds, as well
as the high \mbox{number} of plant species with population trends \mbox{classified}
as ``uncertain''. This is at least partly due to the fact that the
plant monitoring scheme is more recent than the bird one and relies on
fewer monitored squares. These results emphasize the difficulty of
detecting very short-term changes in population dynamics, and the need
to be thoughtful when defining both the sampling effort and focus
period, as stated in \citep{Rigaletal2020,Duchenneetal2022}. It also
suggests that more work need to be done to understand how each
observation contributes to trend estimation and to deal with any
imbalance in sampling efforts across monitoring schemes and over time.
One possible solution could be to weight observations so that each year
contributes equally to the trend, which is the principle behind the
data imputation method offered by the tool TRIM~3 
\citep{PannekoekVanStrien2005}. This idea of weighting observations
depending on a representativity criterion has also been implemented
in well-known indicators such as the Living Planet Index 
\citep{Westveeretal2022}.

Another limitation comes from the data themselves. When trying to
routinely analyze multiple species, we aimed at combining both generic
and relevant model specifications. This was made difficult by the
nature of the collected data, which consists of counts of multiple
individuals at the same time, leading volunteers to round off their
observations to the nearest ten. This is particularly true for the bird
dataset with abundance distributions showing a higher frequency of
tens, thus questioning the choice of the underlying distribution law.
To further assess model adequacy with respect to data structure and
goodness-of-fit, we conducted diagnostic tests using the DHARMa package
\citep{Hartig2024}. We did not detect overdispersion issues for either
plant or bird models, indicating a good fit and the ability to handle
large variance with the chosen beta-binomial and negative binomial
distributions, for plants and birds, respectively. However, we did find
underdispersion in 88 out of 181 plant species, and in approximately
one third of the bird species (56 out of 148). Notably, the strongest
underdispersion occurred in gregarious birds. This is likely due to a
stronger rounding bias in observations (as discussed above), which may
have resulted in lower variances than expected. While underdispersion
leads to fewer concerns than overdispersion, as it tends to produce
more conservative estimates \citep{Bolker2024}, it still warrants
consideration: plant trends are notably affected by relatively large
confidence intervals. Therefore, new functionalities and data
treatments should be explored to address this issue. One solution may
be to use distributions that specifically deal with both
underdispersion and overdispersion, such as the Conway--Maxwell--Poisson
distribution or the generalized Poisson distribution for count data,
and the quasi-binomial for occurrence data. For birds, an additional
step in data processing could also be considered to correct rounding
biases, such as adding random noise to counts. 

\subsection{Applications and next steps for the pipeline}

The proposed pipeline is a valuable tool for systematically documenting
temporal changes in the abundance of common birds or plants. It
provides a proof-of-concept that can be readily adapted to other
monitoring schemes targeting diverse taxa, for which no systematic
appraisal of temporal trends may exist. In France, citizen science
monitoring schemes exist for bats (Vigie-Chiro) and butterflies
(STERF), which are two obvious candidates for such an extension.
Nevertheless, several areas for improvement have been identified.
First, although the pipeline can already handle different
measures/proxies of abundance, namely count and frequency data, it does
not cover biomass data, which may be frequent in insect monitoring 
\citep[see][]{Hallmannetal2017} or percent-cover data, which is common in
plant monitoring. The handling of such data can, however, be easily
implemented in the pipeline, as it requires standard linear mixed
models with a Gaussian distribution. Second, the thresholds used to
classify species into categories of temporal trends, from strong
decline to strong increase, were directly inspired by the IUCN
categories, but may need to be tailored to the specificities of each
taxonomic group. Indeed, IUCN categories may be operational for groups
for which count data are easily collected (large animals such as birds
and mammals), but they may be less relevant for other groups with
different abundance measurements and fewer data, such as plants. This
would reduce the number of species in the ``uncertain'' category and
provide a better vision of the current dynamics of biodiversity across
multiple taxonomic groups. Finally, to make the graphical outputs of
the pipeline accessible to a wide audience, a solution could be to
exploit R functionalities, such as the RShiny package
\citep{Changetal2025}, which is especially adapted to online sharing
with an interactive user interface. By connecting our routine to a web
application, users could update graphics depending on the selection of
species, temporal frames or even relevant thresholds. 

\section*{Acknowledgements}
We wish to thank all the volunteers who collected data for the STOC and
Vigie-flore schemes, without whom these analyses and insights on bird
and plant population dynamics would not have been possible. We also
thank Luc Barbaro for his hypotheses on the population dynamics of the
Eurasian hoopoe.

\section*{Declaration of interests}

The authors do not work for, advise, own shares in, or receive funds
from any organization that would benefit from this article, and have
declared no affiliation other than their research organizations.

\back{}

\section*{Supplementary materials}
Appendixes~1 to~7 and supplementary materials are available on the
journal's website under \printDOI\  or from the author.

\CDRsupplementaryTwotypes{supplementary-material}{\cdrattach{crbiol-184-suppl-Variable-Selection-Procedure.pdf}}
\CDRsupplementaryTwotypes{supplementary-material}{\cdrattach{crbiol-184-suppl-Habitat-Specialization.pdf}}
\CDRsupplementaryTwotypes{supplementary-material}{\cdrattach{crbiol-184-app-1.pdf}}
\CDRsupplementaryTwotypes{supplementary-material}{\cdrattach{crbiol-184-app-2.pdf}}
\CDRsupplementaryTwotypes{supplementary-material}{\cdrattach{crbiol-184-app-3.pdf}}
\CDRsupplementaryTwotypes{supplementary-material}{\cdrattach{crbiol-184-app-4.pdf}}
\CDRsupplementaryTwotypes{supplementary-material}{\cdrattach{crbiol-184-app-5.pdf}}
\CDRsupplementaryTwotypes{supplementary-material}{\cdrattach{crbiol-184-app-6.pdf}}
\CDRsupplementaryTwotypes{supplementary-material}{\cdrattach{crbiol-184-app-7.pdf}}

\printbibliography
\refinput{crbiol20240549-reference.tex}

\end{document}