MeadoWatch: a long-term community-science database of wildflower phenology in Mount Rainier National Park

We present a long-term and high-resolution phenological dataset from 17 wildflower species collected in Mt. Rainier National Park, as part of the MeadoWatch (MW) community science project. Since 2013, 457 unique volunteers and scientists have gathered data on the timing of four key reproductive phenophases (budding, flowering, fruiting, and seeding) in 28 plots over two elevational gradients alongside popular park trails. Trained volunteers (87.2%) and University of  Washington scientists (12.8%) collected data 3–9 times/week during the growing season, using a standardized method. Taxonomic assessments were highly consistent between scientists and volunteers, with high accuracy and specificity across phenophases and species. Sensitivity, on the other hand, was lower than accuracy and specificity, suggesting that a few species might be challenging to reliably identify in community-science projects. Up to date, the MW database includes 42,000+ individual phenological observations from 17 species, between 2013 and 2019. However, MW is a living dataset that will be updated through continued contributions by volunteers, and made available for its use by the wider ecological community. Measurement(s) Timing of four key reproductive phenophases of wildflowers Technology Type(s) Field observations Measurement(s) Timing of four key reproductive phenophases of wildflowers Technology Type(s) Field observations


Background
Increasing temperatures linked to anthropogenic climate change have caused and are expected to continue causing major phenological shifts (i.e., changes in the timing of relevant biological events) of many ecosystems. Higher early season temperatures, changing precipitation patterns, and increasing severity or frequency of extreme weather events have already significantly affected the timing of migration and reproduction across taxa [1][2][3] . These profound phenological changes have reportedly even affected large-scale pheno-patterns, such as homogenizing the leaf-out timing across tree species in the Alps 4 .
Phenology is increasingly recognized as a key indicator of the impact of climate change on natural ecosystems 1 . However, the complexity and species-specific nature of phenological responses requires the use of long-term, multi-species, and high-resolution data to provide assessments of the response to environmental change both at the individual and whole-community level (e.g. 5,6 ). An example of how long-term, open phenological data can contribute to our understanding of ecological change is the Rocky Mountain Biological Laboratory (RMBL) phenology dataset. RMBL researchers have recorded flowering phenology of 135 plant species every other day (recently changed to 3 times per week) since 1973 (DOI:10.17605/OSF.IO/JT4N5). RMBL resolution and scale have allowed researchers to account for factors such as long-term and metric definition variability when studying shifts in community phenology 7 , and interactions between species 8 , in ways that would have been impossible without this long-term dataset. Other large-scale initiatives, such as the National Phenological Network (https://usanpn.org) have expanded the taxonomical and spatial scales of phenological assessments and engaged tens of thousands of volunteers across the United States diverse ecosystems, producing 143 peer-reviewed publications to date. Increasing the number and specially the availability of long-term (2022) 9:151 | https://doi.org/10.1038/s41597-022-01206-8 www.nature.com/scientificdata www.nature.com/scientificdata/ phenological datasets is a remarkable scientific investment to improve our understanding of climate change effects on ecological communities.
However, a challenge to implement and maintain long-term phenological assessments of many species is their costs (monetary and time) (but see e.g. [9][10][11] ). Community-science approaches (also known as citizen-science, public participation in scientific research, or crowd-sourced data) could prove to be a particularly useful tool for acquiring phenological data with remarkable resolution and scale (e.g. 12 ), while simultaneously boosting public participation in science 13 . Furthermore, the increase in scale need not come at the cost of lower data quality, as trained volunteers have consistently shown high accuracy and reliability in locating and determining species and phenophases [14][15][16] . While community-sourced data have its challenges (such as inconsistent temporal resolution 17 ) and requirements for data quality and consistency; it is clear that community science opens a cost-effective and engaging way to enhance traditional scientific approaches 18,19 .
Here, we provide access to a long-term and high-resolution phenological dataset of 17 wildflower species, from one of the most iconic ecosystems in North America: the alpine and subalpine meadows of Mt. Rainier National Park (named Mount Tahoma by the local Salish and Puyallup people, Fig. 1), collected within the MeadoWatch (MW) project (http://www.meadowatch.org/). MW includes data collected by volunteers and scientists on wildflower phenology, soil temperature, and snow coverage on 28 permanent plots located along two elevational transects. Data from the seasons 2020 and 2021 have been (or are being) collected and are in process of data cleaning and preparation. The link provided here will be regularly updated to grant access to the most recent data.

Methods
Study origin and design. The MeadoWatch project (MW) is a project run collaboratively between the University of Washington (UW) and the United States National Park Service to monitor the phenology of alpine and subalpine wildflower species across large elevational gradients in Mount Rainier National Park (Fig. 2). MW was established in 2013 with the goal of understanding long-term effects of climate change on Mount Rainier National Park wildflower communities using community-science approaches. The first MW transect was established along Reflection Lakes, Skyline, and Paradise Glacier trail system in 2013 (9-11 plots). In 2015, MW expanded to include a second transect (15-17 plots) along the Glacier Basin trail (Fig. 1a). The MW transects span around 5 km each, over a 400 m altitudinal gradient (Reflection Lakes: 1490m-1889m a.s.l.; Glacier Basin: 1460m-1831m a.s.l.) Plots are located along the side of each trail, marked with a colored survey marker. The area of each plot is defined by the arm-span of volunteers when they position themselves over the plot marker looking away from the trail (Fig. 2b). While less accurate than marking the corners of plots, this approach was used to avoid establishing permanent structures in wilderness areas within the National Park. The surveyed area in each plot is, on average, 1.25 m 2 . Each plot is also equipped with temperature sensors (HOBO Pendant Logger, Onset Computer Corp.) buried approximately 4 cm below the ground. Sensors are placed at the start of each fall season and removed at the beginning of each summer season for data downloading. The HOBO sensors provide an estimate for the date of snow disappearance and in-situ temperature at 3 hours intervals. Once plots are covered in snow, soil temperatures remain at 0 °C and show no diurnal variation, so that daily changes in temperatures above 1 °C can be used to determine the earliest date without snow cover 20 . We use these approaches to provide dates of www.nature.com/scientificdata www.nature.com/scientificdata/ snow appearance and disappearance, snow cover duration, and minimum soil temperatures for each year and plot. Occasionally, temperature data during the snow disappearing window were lost due to sensor failure or loss of sensors (which occurs because plots are not permanently marked and/or well-meaning visitors remove sensors). This, and the lack of temperature sensors in the first year of the project, resulted in approx. 20% of cases of missing data. In those cases, we used a data imputation method to estimate the missing values based on data from nearby plots and a parallel temperature data collection with 890 total observations. These estimates were highly reliable in filling the data gaps (see Appendix C in 16 for further details).
Due to challenges in species identification, we dropped Canby's licorice-root (Ligusticum canbyi) as a target species in 2016. Consequently, Ligusticum canbyi has limited replication in the database (Fig. 2c). While we included the phenological records of this species for the sake of completeness, we recommend focusing on the other 16 species, which are both better represented (in terms of data coverage) and are free of any potential misidentification issues.

Data collection and volunteer training.
During the summer months, MW volunteers and scientists collect reproductive phenology data with a frequency between 3 and 9 trail reports per week. Each report records the presence or absence of 4 phenophases for each target species present in each of the plots. The phenophases are 'budding', 'flowering', 'ripening fruit', and 'releasing seed'. Phenophases were defined as follows: Budding. The beginning growth of the flower which has not yet opened. A plant is considered budding if buds are present, but the stamen and pistils are not yet visible and available to pollinators.
Flowering. The generally "showy" part of the plant that holds the reproductive parts (stamens and pistils). A plant is considered flowering when at least one flower is open, and the stamens and pistils are visible and available for pollination and reproduction.
Ripening fruit. The fruit develops from the female part of the flower following successful pollination. In the target species, fruits can take many forms, from hard, fleshy capsules, juicy berries, to a feathery tuft on the end of a seed. A plant is in the ripening fruit stage when reproductive parts on at least one reproductive stalk are non-functional and the formation of the fruit part is clearly ongoing (visible), but seeds are not yet fully mature and not yet being dispersed.  www.nature.com/scientificdata www.nature.com/scientificdata/ Releasing seed. After the fruit ripens, seeds are released to be dispersed by gravity, wind, or animals. A plant is considered in the releasing seed stage if seeds are actively being released on at least one reproductive stalk (but there are still seeds present).
A full description, including illustrations for each species' phenophase and identification cues is available in http://www.meadowatch.org/volunteer-resources.html, as well as in Annex 1 -Supplementary Documentation. Multiple phenophases can be present simultaneously, depending on the species, and are noted independently. Additionally, volunteers are also asked to record the presence of snow ('snow covered plot', 'partially covered plot', or 'snow-free plot'), and, since 2017, the presence of damage by herbivory ('presence of browsed stems') on each plot.
In years not impacted by the SARS-Cov-2 pandemic MW volunteers attend an in-person 3-hour botanical and phenological training session taught by UW scientists at the beginning of each sampling season. Volunteers were also provided with detailed species-identification guides, including an extensive description of sampling methods and location of the plots. The trainings for the 2020 and 2021 seasons were held virtually via a series of online training videos. In these years, volunteers took a quiz on wildflower phenology, plant identification and data collection methods after viewing these videos and were required to 'pass' a certain threshold to volunteer (unlimited attempts were allowed). During these virtual trainings, volunteers were provided with digital copies of the species identification guides, with many returning volunteers using printed guides they had kept from previous years.
At the end of their phenological hike, volunteers submit their data sheets either by depositing them in lockboxes located within the park, or by scanning and emailing them directly to mwatch@uw.edu. Data are then entered manually and stored in the UW repositories after being checked for consistency at the end of each sampling season.
The parallel data collection by members of UW's Hille Ris Lambers group (including PI, postdoctoral researchers, graduate students, and trained interns) acted as the following: (i) a quality-control, i.e., allowing us to compare the consistency in phenology assessments between volunteers and scientists, and (ii) a way to increase the temporal resolution and scale of the data, e.g., by reducing early season gaps and 'weekend bias' 17 . This parallel expert sampling was carried out around once a week between 2013 and 2020, showing great consistency between the two groups. For detailed comparisons between volunteers and scientists' assessments see the data validation section (as well as Appendix E in 16 ).
Processed data. In addition to the raw phenological data, we also provide here parameters to construct the year, species, and plot-specific flowering phenology based on the timing of snow disappearance (as in 16 ). Models describe unimodal probability distributions that were fitted with maximum likelihood models to binomial flowering data from each species, year, and plot. These curves have been used to estimate peak flowering dates and diversity and link them to reported visitor experiences 16 . Here, we provide the 3 parameters defining the unimodal curve of flowering probability per species i, plot j and year k: the duration of flowering ( ijk ), the maximum probability of flowering ( ijk ), and peak flowering (in DOY -ρ ijk ); following the equations described in 16 and https://github.com/ajijohn/MeadoWatch).
The parameters of these probability distribution curves are ready-to-use values that can be broadly and easily used to estimate floral compositional change over past seasons due to changing environmental conditions-for example, to inform plant-pollinator interaction networks if combined with pollinator behavioral data (e.g. 21 ).

Data Records
MW data are stored permanently in the repository: https://doi.org/10.5061/dryad.g1jwstqs2 22 , as well as privately in UW's private repositories, to ensure public availability and future safety of the data.
When MW is updated, files will be updated and named accordingly in both repositories. The DOI for MW data, however, will remain unchanged and constantly updated using version control.
MW dataset consists of 6 files in csv format: 1) MW_metadata.xlsx: this file contains information on each variable, including variable name, description, units, file, reference for methods. It also includes the equivalency between 4-letter species code used in the reports and the full botanical name. -Bud_rank: each phenophase for each species at each plot was also ranked based on relative abundance, from 1 (most common phenophase) to 4 (least common phenophase) observed. These measurements occurred from 2013 to 2015 but were dropped in subsequent assessments (replaced with NA in the database subsequently). -Training_type: whether the observer has been trained in-person (noted as 'Traditional') or only on-line (noted as 'Online'). Observers that were attended online trainings after attending in-person ones were considered as having a traditional training.

technical Validation
Temporal resolution. Phenological reports span the growing season (typically between May and October), but have consistently good temporal resolution between June and September (Fig. 1c,d). The average time between reports across the whole growing season was variable from year to year but was consistently between 3 and 5 assessments per week for each of the trails (Table 1). www.nature.com/scientificdata www.nature.com/scientificdata/ comparison of scientist and volunteer phenological assessments. To estimate the consistency in phenological assessments between scientists and volunteers, we compared scientist and volunteer reports collected on the same day and trail (Fig. 3). After pairing the reports, we followed a confusion matrix approach; where 'true positives' and 'true negatives' would be instances of agreement between scientist and volunteers, respectively; 'false positives' would be instances when scientists recorded absence of a phenophase and volunteers recorded presence; and 'false negatives' are the opposite case (volunteer reported presence while scientist recorded absence). We are aware that many of our volunteers are highly botanically trained (even more than some of our scientists!), therefore this terminology and approach has the sole objective of quantifying the agreement or disagreements in species and phenophases ID.

Glacier Basin Reflection Lakes
Using this framework, we calculated the accuracy, sensitivity, and specificity metrics for each trail, phenophase, and species as follows:  Therefore, within our study: -accuracy represents the percentage of consistent assessments (either positives or negatives); -sensitivity indicates the chance for a phenological observation to be only captured by volunteers, but not by professional scientists; -and specificity the chance for a phenological observation to be only captured by a scientist, but not by a volunteer.
We found consistently high accuracy for the assessments of all phenophases (>82%) and species (>73%, Figs. 3, 4). As expected, the flowering phenophases showed the highest accuracy, specificity, and sensitivity. However, differences between phenophases were not significant (Fig. 4a). Similarly, differences between transects were small and not consistent. While accuracy and specificity were consistently high across all species (Fig. 4b,d), sensitivity was lower and more variable among species (Fig. 4c). Sensitivity was particularly low for the seeding and fruiting phenophases, indicating a higher chance for volunteers to miss the fruiting or seeding bodies of our target species. By species, Mertensia paniculata and Castilleja parviflora had the lowest accuracy www.nature.com/scientificdata www.nature.com/scientificdata/ and specificity. This may be caused by their smaller size and the more subtle transition between similar-looking phenophases (e.g., fruiting and releasing seeds stages in Castilleja parviflora). Note, however that in both cases accuracies and specificities are still >80% for both species (Fig. 4b,d). The species with lowest sensitivity (<60%) were Microseris alpestris, Angelica arguta, and Aster ledophyllus, which were particularly challenging to detect for volunteers in their seeding state (Fig. 4c).
All the analyses were performed in R 23 . Graphs were exported from R to Inkscape Software (https://inkscape.org/) to produce the final published figures.

Data exploration tool.
To interactively showcase the possibilities of the collected data and allow users to easily explore the data, we developed a Shiny App 24 . Shiny App is an interactive web application framework written in R 23 that aims at creating cross-platform web experiences with customizable user-interface (UI) components. Shiny Apps can be launched via any locally installed R environment or can be hosted on the shinnyapps webserver (shinyapps.io). The data exploration tool is accessible at https://explorations.shinyapps.io/ MeadoWatch/. Using dropdown menus, users can search and explore the raw data ('Phenology at sites' menu, Fig. 5a), the species-specific phenological curves ('Model fitting curves' , Fig. 5b), the number of flowering species for each day of the year in each of the projects year ('Flowering richness'), and the connection between flower phenology and hiker satisfaction reported in 16 ('MW WTA analysis'). Data releases and improving functionalities will be progressively implemented to this application and announced on the GitHub and 'Releases' section of the web.

Usage Notes
Data availability and use. This paper represents MW's first data release, which will be updated regularly via the provided links. MW data is freely available to use for any non-commercial purposes by citing this reference or the DOI to the database 22 . Commercial use of the database requires express authorization by the MW coordinator (Prof. Dr. Hille Ris Lambers). For privacy, the identities of MW volunteers in the database have been anonymized.
Applications. MW data can be used to explore changes in wildflower community phenology, species climate sensitivity, and could be combined with other datasets, such as pollinator data, to study shifts in trophic interactions in North American natural ecosystems. The database has been used to date to study the reliability of wildflower phenology estimates using traditional and community science methods 25 , to show social-ecological Fig. 4 Agreement between scientists and volunteers in the estimation of different phenophases (a) and across species (b,c,d). Mean value per phenophase or species are shown for each group. Accuracy (b) represents the percentage of consistent assessments (either positives or negatives), (i.e., 80% accuracy would indicate that 80% of reports show agreement between scientists and volunteers). Sensitivity (c) represents the chance for a phenological observation to be captured only by volunteers but not by scientists (i.e., 80% sensitivity would indicate that 20% of reports by volunteers missed that species). Finally, specificity (d) represents the chances for a phenological observation to be captured only by scientists but not by volunteers (i.e., 80% specificity would indicate that 20% of reports by scientists missed that species). Species are sorted in decreasing order. Averages and standard errors are shown per species. Species significantly better (green) or worse (orange) than the across species average (vertical dotted line) are highlighted.
www.nature.com/scientificdata www.nature.com/scientificdata/ mismatches in changing phenology 12 , as 'ground truth' to fit models that detect montane phenology from satellite imagery 26 , and to explore the effects of warming climate on flowering diversity and timing and how it affects the experience of visitors to National Parks 16 . We also provide information on volunteer identity (anonymized) and training type, which can be used to evaluate methods characteristics of community-science volunteers. Further information on MW volunteers may be available upon private request to the authors. We hope that by making these data openly and easily accessible, multidisciplinary works and models can help better understand phenological changes in natural ecosystems.

code availability
The most up-to-date data will be available and updated regularly at: https://doi.org/10.5061/dryad.g1jwstqs2. The code to replicate all the analyses, as well as the raw data finalized as of August 2021 is openly available via: https:// github.com/rdmanzanedo/MeadoWatch_SDATA_2022.git. The code to launch the Shiny app data exploration tool is available in the same GitHub folder and as a browser application here: https://explorations.shinyapps.io/ MeadoWatch/.

Fig. 5
Screen captures of the interface of the data exploration tool (https://explorations.shinyapps.io/ MeadoWatch/). The raw data exploration menu and search options (a) and multi-year curves of flowering probability for Erigenon peregrinus (ERPE) (b) are shown. The code to run the data exploration tool locally is also available in the MW GitHub repository (https://github.com/rdmanzanedo/MeadoWatch_SDATA_2022.git).