eBird Status and Trends Data Products Changelog

2022 Changelog

Data Version: 2022 (available November 2023)

Citation:
Fink, D., T. Auer, A. Johnston, M. Strimas-Mackey, S. Ligocki, O. Robinson, W. Hochachka, L. Jaromczyk, C. Crowley, K. Dunham, A. Stillman, I. Davies, A. Rodewald, V. Ruiz-Gutierrez, C. Wood. 2023. eBird Status and Trends, Data Version: 2022; Released: 2023. Cornell Lab of Ornithology, Ithaca, New York. https://doi.org/10.2173/ebirdst.2022

Status

Data Inputs

eBird Checklists
  • CHANGED: Checklists are included for January 1 2008 through December 31 2022, updated from January 1 2007 through December 31 2021.
  • CHANGED: Checklists up to 30 km in length are now included for species that are run for the ocean only (e.g., Northern Gannet).
  • CHANGED: Checklists with duration less than 0.0167 hours (1 minute) are dropped. These are primarily checklists with incorrect duration information.
  • CHANGED: Checklist centroids derived from tracks are now calculated using great circle distance, not sinusoidal distance.
  • CHANGED: The maximum allowable number of observers on a checklist is 50. All other checklists are dropped.
Effort Covariates
  • ADDED: Rate (as kilometers per hour) has been added as an effort covariate. Prediction is made to a value of 2 kmph, as the units for duration is 1 hour and distance is 2 kilometers. For the range boundary prediction, rate is set as the result of effort_hrs and effort_distance_km having been maximized with partial dependence values separately.
  • FIXED: Rainfall and Snowfall had a bug that resulted in them often being all 0s, due to precision errors. This has been corrected and tested.
  • CHANGED: CCI Summary: The main changes to the calculation of CCI are in
  1. What kind of model is fit to checklist species richness using predictive features, and
  2. How deviations in model predictions are attributed to particular observers and checklists.

Details The foundation of CCI is a predictive model of checklist-level species richness (S; i.e. number of species). In updating CCI, changes were made to both the form of the predictive model of S and to the method that attributes variation in richness to particular observers.

Prior to Version 2022, predictive features comprised weather, landcover, habitat diversity, protocol, day of year, and variables that are particular to the observer: observer_id and checklist_number (i.e., index of how many checklists a user has ever submitted from any stixel to eBird; not to be confused with checklist_id). A mixed-effects generalized additive model (GAM) was fit to S. This GAM used as predictive features the natural log of checklist_number, a smooth spline of solar_noon_diff, and the raw values of all other predictors, with a random effect specification for observer_id and checklist_number. The model was used to make predictions pi of S to data representing a “standardized search”, in which all features except observer_id and checklist_number were held constant (at the column-wise mean) across observations. CCI was derived from the variation in resulting predictions, and scaled to have mean 0 and variance 1.

$$ CCI_{i} = \\(pi - mean(p)\\) / sd(p) $$

Version 2022 changed the functional form of the predictive model from a (mostly) linear mixed-effects model to a random forest. Further, it removed observer_id and checklist_number from the suite of predictive features; the model is now blind to person-specific effects. Instead, predictions to real data absent any personal information establish conditional expectations of richness given habitat, effort, weather, etc. Each expected value parameterizes a Poisson distribution, which is used to compute the exceedance probability of the actually-observed S, which is then mapped to a standard-normal quantile. A GAM with a “factor smooth” basis for checklist_number and observer_id is applied to smooth the raw values for each observer. CCI currently comprises these smoothed values.

Environmental Covariates
  • CHANGED: Covariate assignment now uses proper circular buffers for neighborhood calculations (with 1.5 km radius) instead of the previously used sinusoidal buffer, which had many locations of high skew across the globe.
  • ADDED: Information about the moon has been added using two covariates from the R suncalc package. Moon fraction represents the fraction of the disk of the moon that was illuminated at a given time and location and Moon altitude represents the altitude of the moon (above or below the horizon, in radians) at a given location and time.
  • ADDED: The Joint Research Center Global Surface Water data has been added as yearly variables, representing the binary presence of either seasonal or permanent water (JRC/GSW1_4/YearlyHistory), calculated as percent land cover and edge density within a neighborhood. This dataset is 30 m in spatial resolution.
  • ADDED: Elevation data at 30 m resolution has been added from the ASTER Global Digital Elevation Model. This is represented as mean and standard deviation within neighborhoods.
  • ADDED: MODIS 16-day Enhanced Vegetation Index (EVI) has been added from MOD13Q1. This has been summarized as mean and standard deviation within neighborhoods. As this dataset is not available for water and has an artificial boundary at the northern and southern latitudes based on availability of light, we have added a boolean covariate has_evi that describes whether the covariate was available at a given date and location.
  • ADDED: Data describing shorelines from Sayre et al. 2021 has been included. This includes means and standard deviations for: wave height, tidal range, chlorophyll, turbidity, sinuosity, slope, and outflow density; class densities (as km of coast per square km of area in neighborhood) for for four classes of erodibility, and class densities (as km of coast per square km of area in neighborhood) for 23 Ecological Marine Units (EMUs) that describe the sea surface temperature, salinity, and dissolved oxygen, as well as covariates describing the unique number of erodibility and EMU classes in each neighborhood. As with EVI, we have included a boolean has_shoreline covariate, as the shoreline covariates are not spatially exhaustive, describing whether the covariate was available at a given location.
  • UPDATED: MCD12Q1 LCCS land cover, land use, and hydrology data were updated to version 6.1.

Workflow and Code Changes

Base Model
  • CHANGED: Both migrants and residents use circularized time as the covariates for day of year, as sine and cosine of day normalized within a stixel.
  • CHANGED: The binary presence/absence occurrence threshold at the base model level now uses mccf1, replacing Cohen’s Kappa.
Prediction
  • CHANGED: The calendar dates for grid prediction have changed as a result of converting the prediction values to integers (formerly they included fractional day information).
  • CHANGED: The prediction values for both effort_distance_km and effort_hrs are set to their 90th quantiles when making predictions to determine the range boundary. Previously these were chosen to maximize the partial dependence (PD) curve.
  • CHANGED: The prediction values for CCI and time of day (solar_noon_diff) are now chosen to maximize the abundance partial dependence (PD) constrained to values where the species was detected. Previously they were chosen using the occurrence partial dependence curve and were not constrained to detections.
  • CHANGED: When maximizing the prediction value for CCI at a stixel level, the allowable range of values is now 0-2, up from 0-1.85, based on the range of the new version of CCI values overall.
  • CHANGED: The prediction value for effort_distance_km is now 2 km, to more closely reflect the distribution of checklists and to increase overall signal.
  • CHANGED: weather optimization arise now done for relative abundance, not occurrence.
  • CHANGED: PD maximization of solar_noon_diff allows the full range of quantiles to allow selection of the highest and lowest quantile values which are often nocturnal. Previously the outermost quantile values were not allowed for selection.
Ensemble
  • CHANGED: With the replacement of the base model binary presence/absence occurrence threshold with MCC-F1, the ensemble level percent above threshold (PAT) cutoff value has been fixed at 0.14 (interpreted as a species being found at least once a week).
  • ADDED: In the ensemble support calculation, a new product has been added, spatial coverage. This represents the fraction of 3 km grid cell-weeks that have checklists within a given stixel, averaged across the ensemble (essentially a spatial smooth. This weekly layer is then used to mask all predictions from all species values where the spatial coverage value is below 0.00025. This helps control extrapolation in places like Russia and central Africa.
  • CHANGED: The ensemble support site selection probability is now 0 for unsampled islands.
  • CHANGED: The ensemble support-based site selection probability mask has been changed to 0.0025. Ocean run species no longer use the site selection probability mask.
Data Products
  • UPDATED: The prediction definition is now: the {occurrence, count, or relative abundance} of individuals of a given species detected by an expert eBirder on a 1 hour, 2 kilometer traveling checklist at the optimal time of day. Predictions have been optimized for user skill, hourly weather and moon conditions, specific for the given region, season, and species, in order to maximize detection rates.
  • REMOVED: Partial dependence values are no longer calculated or distributed.

2021 Changelog

Data Version: 2021 (available November 2022)

Citation:
Fink, D., T. Auer, A. Johnston, M. Strimas-Mackey, S. Ligocki, O. Robinson, W. Hochachka, L. Jaromczyk, A. Rodewald, C. Wood, I. Davies, A. Spencer. 2022. eBird Status and Trends, Data Version: 2021; Released: 2022. Cornell Lab of Ornithology, Ithaca, New York. https://doi.org/10.2173/ebirdst.2021

Data Inputs

eBird Checklists

  • CHANGED: checklists are included for January 1 2007 through December 31 2021, updated from January 1 2006 through December 31 2020.
  • CHANGED: Observations reported as escapees under the new eBird exotic species protocols are excluded from analysis.

Environmental Covariates

  • UPDATED: Data for 2020 was added for the primary land cover data source, MCD12Q1.

Workflow and Code Changes

General

  • ADDED: Prediction grid locations for the ocean are now available as a choice to model a species as land or water.

Spatiotemporal Partitioning

  • CHANGED: The adaptive partitioning algorithm (AdaSTEM) now grid samples the training data before stixels are defined.
  • CHANGED: The projection initialization of each stixel iteration is now fully randomized, previously it was constrained to keep boundaries in the ocean.
  • CHANGED: Stixels are now allowed to recurse one size smaller, to approximately 90km on a side, and remain one size larger (3000km on a side), except for resident-specific stixels where the maximum remains 1500km on a side, for computational reasons.
  • CHANGED: There is now a separate AdaSTEM partitioning for residents that uses the full year of data instead of a 28 day window. The training data for these partitions are also grid sampled before definition. The stixel parameters are set to have a maximum of 65,000 checklists per stixel over the full year, after grid sampling, and a minimum of 6,500 checklists per stixel (e.g., stixels are not allowed to be subdivided if they contain less than this amount).

Model Ensemble

  • CHANGED: Models are now run for 200 replicates (folds).
  • CHANGED: The percent above threshold (PAT) cutoff has been replaced with a data-driven maximization of the MCC-F1 curve (https://arxiv.org/abs/2006.11278), constrained between 0.05 and 0.25. The training data are grid sampled before optimizing using the MCC-F1 curve and 25 realizations are done before taking the median PAT value. For migrants, this is done weekly, for residents across the whole year.
  • CHANGED: The process for selecting the ensemble support cutoff or threshold (the number of models required to show predictions) has been updated to have the training data grid sampled first, then optimized for a true positive rate of 99%, with the cutoff constrained between 0.5 and 0.9. For migrants, this is done weekly, for residents across the whole year. This process is done 25 times and then a median threshold value is selected.
  • CHANGED: The site selection probability layer has been significantly improved. In the binary classification model, prediction grid locations that are >= 50% overlapped by a 1.5km buffer of checklist locations have been removed. This resolves the previous, erroneously low values in dense, urban areas and more accurately reflects the true probability of site selection in these areas. This change only impacts species estimates in places with a site selection probability value of less than 0.5%, where species estimates are masked.

Base Model

  • CHANGED: The grid sample method now retains all unique values of factor variables (e.g., island).
  • CHANGED: The grid sampler oversamples detections to achieve 25% detection probability in the training dataset. Previously the grid sampler would often overshoot the 25% target and excessively duplicate detections. This has been corrected so that oversampling never yields detection probabilities greater than 25% and detections are duplicated at most 25 times.
  • CHANGED: Mean spatial coverage of each stixel is now correctly estimated as the proportion of 3 km pixels that contain checklists.

Fit and Predict

  • CHANGED: Maximization of partial dependencies for prediction (e.g., CCI) no longer allows selection of the highest and lowest extreme quantile values, to prevent extrapolation.

Residents

  • CHANGED: Along with a resident-specific AdaSTEM partitioning, resident models now predict all weeks of the year in a single stixel. Previously, resident models used data from the whole year for training, but only predicted the four weeks in a stixel, similar to the way migrants are modeled.

Data Products

  • CHANGED: The occurrence model prediction values for effort variables are now set at 1 hour and 1 kilometer. Previously, the effort variable values used for the occurrence model prediction were the same as those used for the occurrence model, which sought to maximize detection by optimizing the distance and duration effort variables to capture as much signal as possible, up to 12 hours (6 hours in this version) and 10 kilometers. These prediction values are retained for the presence/absence estimation.
  • CHANGED: The prediction value for Checklist Calibration Index (CCI) is now maximized within each stixel using the partial dependencies. Previously, the value for was set at a fixed value of 1.85 for all species and stixels.
  • CHANGED: Partial dependencies are now only generated for the first 50 folds, to reduce computational cost.
  • CHANGED: To show “year-round” on a seasonal map now requires only 0.1% overlap between breeding and non-breeding seasons. Previously, all four seasons and an overlap of greater than 5% was required.
  • REMOVED: Habitat plots and numerical summaries have been removed from the website.

2020 Changelog

Data Version: 2020 (available Fall 2021)

Citation:
Fink, D., T. Auer, A. Johnston, M. Strimas-Mackey, O. Robinson, S. Ligocki, W. Hochachka, L. Jaromczyk, C. Wood, I. Davies, M. Iliff, L. Seitz. 2021. eBird Status and Trends, Data Version: 2020; Released: 2021. Cornell Lab of Ornithology, Ithaca, New York. https://doi.org/10.2173/ebirdst.2020

Data Inputs

eBird Checklists

  • CHANGED: checklists are included for January 1 2006 through December 31 2020, updated from January 1 2005 through April 15 2020.
  • CHANGED: all species now use all data globally and are not run for spatial subsets. Previously, primarily Western Hemisphere species were run only for that spatial extent.
  • CHANGED: checklists using the Stationary protocol now include tracks and are used as long as the distance of the track for this protocol type is less than 700 meters.
  • CHANGED: The spatial location for checklists at eBird Hotspots has been changed from the user-reported location to the centroid of all tracks associated with the hotspot.
  • FIXED: Previously, some historical checklists that lacked complete effort information had been included. These have now been excluded.

Environmental Covariates

  • CHANGED: SRTM15+ ~250m elevation and bathymetry replaces the ~1 kilometer SRTM30+ elevation and bathymetry product.
  • CHANGED: The single year of Nighttime Lights has been replaced with by-year assignment for 2014-2020 using the EOG Annual VNL v2 product.
  • CHANGED: The Global Intertidal Change dataset has been updated to version 1.2 which includes a new three-year time step covering 2017 through 2019.
  • CHANGED: Continents now have unique identifiers in the island categorization. Previously, all continents were treated as the same “mainland” value.
  • ADDED: Hourly weather variables have been assigned at 30 kilometer spatial resolution using the Copernicus ERA5 reanalysis product.
  • ADDED: 90m eastness and northness (combined slope and aspect) topographic variables from Amatulli et al. 2020 are included in addition to 1 kilometer eastness and northness.
  • FIXED: Source data updated for 2017-2019 for MCD12Q1 which had reported classification errors.

Workflow and Code Changes

Spatiotemporal Partitioning

  • CHANGED: The adaptive partitioning algorithm (AdaSTEM) now uses an Icosahedron Gnomic projection that generates partitions with largely conformal stixel boundaries across the globe.
  • CHANGED: The temporal width of AdaSTEM partitions has been changed from 30.5 days to 28 days.

Model Ensemble

  • CHANGED: The percent above threshold (PAT) cutoff for 3km grid cells to be reported as present has changed from 0.1 to 0.143, to accommodate increased occurrence rates as a result of including hourly weather to account for variation in detection rates.

Resident Methodology

  • CHANGED: Residents now have a suite of independent settings designed for species with strong spatiotemporal stationarity. These include the following:
    • Each stixel loads the full year of training and test data, not just the 28 day window associated with the given stixel.
    • The DAY predictor is encoding cyclically using sin and cosin transformation to allow the model to wrap the year.
    • The spatiotemporal grid sampling now seeks a maximum sample size of 65,000 checklists in a given stixel (for migrants this value is 5,000).

Data Products

  • CHANGED: The count model prediction values for effort variables are now set at 1 hour and 1 kilometer. Previously, the effort variables used for the count model prediction were the same as those used for the occurrence model, which sought to maximize detection by optimizing the distance and duration effort variables to capture as much signal as possible, up to 12 hours (6 hours in this version) and 10 kilometers.
  • CHANGED: Zeroes in data products that are outside of the prediction area for species (also known as assumed zeroes) now require, on average, across the up-to 100 models in the ensemble, 0.5% of 3km grid cells filled with at least 1 checklist for a given week to be reported as zero. Previously, this was 0.1% of 3km grid cells. This has been adjusted to offer a more appropriately conservative representation of where absence can be assumed based on overall data volume.
  • ADDED: Locations (3km grid cells) with less than a 0.5% mean site selection probability are now masked out of the final data products and reported as NA. Mean site selection probability is calculated weekly in a species-agnostic AdaSTEM workflow that estimates the probability that a location of a given habitat configuration will be visited in a given region and season.
  • ADDED: Spatial representations of predictive performance metrics and other individual model-level summaries are being generated as 27km GeoTIFFs for each week of the year. The spatialization is done by assigning the stixel-level values to every 27km grid cell within the stixel and then averaging across stixels to determine regional metrics.
  • FIXED: The Caspian Sea is now masked out of all data products.
  • CHANGED: Raw test data that does not receive model predictions has been removed from the calculation of predictive performance metrics. Previously, this type of test data was used as a form of assumed absence in the calculation of binary predictive performance metrics.
  • ADDED: Predictions to 3km grid cells now include a standardization of hourly weather within each individual model. The hourly weather values set for prediction are based on a maximization of occurrence estimates between the 80th and 90th percentiles.
  • CHANGED: Calculation of individual model partial dependencies now uses train out of bag data. Previously, train in bag data was used.
  • ADDED: Predictor Importance and Partial Dependency products are now included for both occurrence rate and count models. Previously, these products were only available for the occurrence rate model.
  • CHANGED: The time covariate used in the models, calculated as the difference between the local checklist time and solar noon at the checklist location, has been changed to use the temporal midpoint of the checklist for the calculation. Previously, the time at the start of the checklist had been used for this calculation.
  • FIXED: The temporal centroid of individual models, used with predictor importance and partial dependencies, has been changed to represent the mean date of train in bag data. Previously, this was a mean of all train, test, and all four weeks of 3km grid cell location data.
  • CHANGED: Regional habitat association charts are based on a weighted summary of stixel-level predictor importance and partial dependence estimates, with the weighting determined by the proportion of the region covered by each stixel. Previously, stixel centroids were used to determine the set of stixels contributing to a given region, with crude approximations of the stixels as rectangles in lat-lon coordinates being used to determine the overlap-based weighting. Now, the exact stixel shape is used when calculating regional habitat associations, by considering the exact set of 27km grid cells falling within each stixel, to determine both the set of stixels used in habitat summarization and the overlap-based weighting for a given region.
  • CHANGED: Habitat and regional abundance and range statistical summaries are now computed for all species, globally, using the Natural Earth Data Admin 1 data for summarization.

Expert Review

  • CHANGED: Animations are no longer being reviewed for resident species.

2019 Changelog

Data Version: 2019 (available Fall 2020)

Citation:
Fink, D., T. Auer, A. Johnston, M. Strimas-Mackey, O. Robinson, S. Ligocki, W. Hochachka, C. Wood, I. Davies, M. Iliff, L. Seitz. 2020. eBird Status and Trends, Data Version: 2019; Released: 2020. Cornell Lab of Ornithology, Ithaca, New York. https://doi.org/10.2173/ebirdst.2019

Data Inputs

eBird Checklists

  • CHANGED: Checklists are included for January 1, 2005 through April 15, 2020, updated from January 1, 2014 through December 31, 2018.
  • ADDED: Include checklists from the International Shorebird Survey (ISS) as complete for shorebird species.
  • CHANGED: Checklists where “slashes” (representing two similar species) are non-zero now have child species set to “X” (present-only, no count info).
  • FIXED: Subspecies did not always roll up to species-level correctly.

Environmental Covariates

  • CHANGED: ASTER Global Water Bodies Database at 30m for ocean, river, and lakes replaces MOD44W which was 500m resolution, had only land or water classification, and only ran through 2015.
  • ADDED: GLOBIO Global Roads Inventory Project (GRIP) as road density (m/km2) for five classes of roads.

Workflow and Code Changes

Spatiotemporal Partitioning

  • CHANGED: The adaptive partitioning algorithm (AdaSTEM) now uses projected coordinates (sinusoidal) and meters instead of unprojected coordinates and degrees.
  • CHANGED: AdaSTEM partitions are now 1500 kilometers on a side at their largest and 187 kilometers on a side at their smallest.
  • CHANGED: AdaSTEM rules now split partitions if they contain more than 16,000 checklists or are larger than 1500 kilometers on a side.
  • CHANGED: AdaSTEM now reverts individual partitions back to the next largest size if any of the partition children contain less than 500 checklists and are not mostly open water. Partitions are never allowed to revert back to partitions that are 1500 kilometers or more on a side.

Model Ensemble

  • ADDED: Individual models now report 0 for predictions if the training data set contains less than 10 positive observations of a species and the mean spatial coverage within the model is greater than or equal to 5%.
  • CHANGED: Range boundaries are now set weekly to have the highest level of ensemble support, between 50% and 95% of models, while including at least 99.5% of positive observations, changed from being fixed at 75% of models in previous versions.
  • CHANGED: Zeroes in data products that are outside of the prediction area for species (also known as assumed zeroes) are now based on the mean spatial coverage of checklists within those areas. For locations where species-specific models did not report zero or non-zero predictions, locations need to have, on average, across the up-to 100 models in the ensemble, 0.1% of 3km grid cells filled with at least 1 checklist for a given week to be reported as zero. Previously, these locations required 95% of models at a given location to have had at least 50 complete checklists for the given week.

Seasonal Products

  • ADDED: When averaging weekly estimates to represent resident species, reviewers select a subset of weeks, as opposed to having previously averaged the entire year.

Data Products

  • ADDED: There are now 184 species modeled at a fully global extent. The overall species total is now 807.

Expert Review

  • ADDED: Expert reviewers now assign quality scores for the full-year, animations, and all seasons.