Site selection
We selected 104 sites covering a range of ecological, land cover, and climate conditions across North America (Table 1). These sites were selected because they are part of either the National Ecological Observatory Network (NEON) or AmeriFlux network, all have PhenoCams, and each has at least one year of available flux data between 2017 and 2021. Among the included sites, 44 are part of the NEON.
PlanetScope image database compilation
The LSP metrics included in the dataset are derived from a database of daily 3 m PlanetScope imagery. To compile this database, a Python script was created to search, request, and download imagery using Planet’s RESTful API interface (https://developers.planet.com/docs/apis/data/). For each site, the area of interest (AOI) was defined using a GeoJSON file that prescribed a 10 by 10 km box centered over the flux tower at each site. Each GeoJSON was then used to submit search requests to the API. As part of the search process, the following filters were applied to ensure that good quality images with adequate clear sky views and high-accuracy geolocation were downloaded: (1) quality category identified as ‘standard’; (2) cloud cover less than or equal to 50%; and (3) ground control is ‘true’. Filtering was performed using all available PlanetScope ‘PSScene4Band’ imagery from 2016 to 2022. Once the API completed the search, the Python script read the search results, submitted orders, and the selected imagery was downloaded from Planet’s cloud-based system to local storage. During execution of the Python script, a log file was created to keep track of successful and failed orders. If an order failed, the script was run again targeting the specific order that failed. The resulting dataset included over 1.8 million unique files with, on average, 3,885 scene images for each site (i.e., the number of images, on average, that overlap part of each 10 by 10 km site), and had a total volume of 62.2 TB.
Image processing
To ensure that high-quality image time series were used to generate LSP metrics, we used PlanetScope per-pixel quality assurance information to exclude pixels that had low quality in all 4 bands (i.e., blue, green, red, and near-infrared). Specifically, we excluded pixels where the Unusable Data Mask (layer ‘umd’) was not 0 (i.e., we retained pixels that were not cloud contaminated or located in non-image areas) and pixels where the Usable Data Mask (layer ‘umd2’) is 0 (i.e., we retained pixels that were not contaminated by snow, shadow, haze, or clouds). We then cropped all the images to exclude pixels outside of the 10 by 10 km window centered over each tower. We selected this window size based on published results showing that 80% of the average monthly footprint at eddy covariance towers ranges from 103 to 107 square meters22. Note that the swath for PlanetScope imagery often did not cover entire sites and some sites (e.g., the tall tower at US-Pfa) have larger footprints than other sites. Similarly, most sites had multiple PlanetScope image acquisitions on the same day. To create image time series, we mosaiced all available imagery at each site on each date, and, under the assumption that geolocation error was non-systematic and modest, we created a single image for each date using the mean surface reflectance for pixels with multiple values on the same day. The resulting database of daily surface reflectance images were sorted in chronological order, sub-divided into 200 sub-areas at each site (i.e., 0.5 km2 each), and then stored as image stacks to facilitate parallel processing to estimate LSP metrics, where each image stack included all images from July 1, 2016 through January 31, 2022.
Creation of daily EVI2 time series
To estimate LSP metrics we adapted the algorithm described by Bolton et al.19, which was originally implemented to estimate LSP metrics from harmonized Landsat and Sentinel-2 (HLS) imagery, for use with PlanetScope imagery. Prior to LSP estimation, daily images of the two-band Enhanced Vegetation Index30 (EVI2) data were generated from PlanetScope imagery and then interpolated to create smooth time series of daily EVI2 values at each pixel in three main steps. First, sources of variation related to clouds, atmospheric aerosols, and snow were detected and removed from the EVI2 time series at each pixel based on data masks provided with PlanetScope imagery (described above) and outlier detection criteria (i.e., de-spiking and removing negative EVI2 values). Second, we identified the ‘background’ EVI2 value (the minimum EVI2 value outside of the growing season) based on the 10th percentile of snow-free EVI2 values at each pixel. Any dates with EVI2 values below the background value were replaced with the background EVI2. Third, penalized cubic smoothing splines were used to gap-fill and smooth the data to create daily EVI2 time series across all years of available data. Complete details on these steps are given in Bolton et al.19. This approach has been tested and shown to yield PlanetScope EVI2 time series that are consistent with both EVI2 time series from HLS imagery and time series of the Green Chromatic Coordinate (GCC) from PhenoCam imagery26. We used the EVI2 instead of other vegetation indices such as the Enhanced Vegetation Index (EVI) or the Normalized Difference Vegetation Index (NDVI) because EVI2 is less sensitive to noise from atmospheric effects relative to EVI and is less prone to saturation over dense canopies and noise from variation in soil background reflectance over sparse canopies relative to the NDVI30,32. Thus, phenological metrics from EVI2 time series tend to have better agreement with PhenoCam observations than corresponding metrics from NDVI33.
Identifying phenological cycles
Prior to estimating LSP metrics, we first identity unique growth cycles by searching the period before and after each local peak in the daily PlanetScope EVI2 time series. To be considered a valid growth cycle, the difference in EVI2 between the local minimum and maximum was required to be at least 0.1 and greater than 35% of the total range in EVI2 over the 24-month period centered on the target year ± 6 months. The start of each growth cycle is restricted to occur within 185 days before the peak of the cycle and at least 30 days after the previous peak. The same procedure was applied in reverse at the end of the cycle to constrain the range of end dates for each growth cycle. This procedure is applied recursively over the time series until each local peak has been assessed and all growth cycles (with associated green-up period, peak greenness, and green-down period) are identified in the time series at each pixel. As part of this process, the algorithm provides the number of growth cycles identified for each year in the time series.
Retrieving LSP metrics
LSP metrics are estimated for each pixel in up to two growth cycles in each year. If no growth cycles are detected, the algorithm returns fill values for all timing metrics, but does report values for the four annual metrics: EVImax, EVIamp, EVIarea, and numObs (see below). If more than two growth cycles are detected, which is rare but does occur (e.g., alfalfa, which is harvested and regrows multiple times in a year), the algorithm records 7 LSP metrics for each of the two growth cycles with the largest EVI2 amplitudes. The resulting dataset includes seven ‘timing’ metrics that identify the timing of greenup onset, mid-greenup, maturity, peak EVI2, greendown onset, mid-greendown, and dormancy. These metrics record the day of year (DOY) when the EVI2 time series exceeds 15%, 50%, and 90% of EVI2 amplitude during the greenup phase, reaches its maximum, and goes below 90%, 50%, and 15% of EVI2 amplitude during the greendown phase. In addition, the algorithm records three complementary metrics that characterize the magnitude of seasonality and total ‘greenness’ at each pixel in each growth cycle: the EVI2 amplitude, the maximum EVI2, and the growing season integral of EVI2, which is calculated as the sum of daily EVI2 values between the growth cycle start- and end-dates (i.e., from greenup onset to dormancy).
Quality assurance flags
Quality Assurance (QA) values are estimated at each pixel based on the density of observations and the quality of spline fits during each phenophase of the growing season. A QA value of 1 (high quality) is assigned if the correlation between observed versus fitted daily EVI2 values is greater than 0.75 and the maximum gap during each phase is less than 30 days. A QA value of 2 (moderate quality) is assigned if the correlation coefficient is less than 0.75 or the length of the maximum gap over the segment is greater than 30 days. A QA value of 3 (low quality) is assigned if the correlation coefficient is less than 0.75 and the length of the maximum gap over the segment is greater than 30 days. A QA value of 4 is assigned if no growth cycles were detected or insufficient data were available to run the algorithm.
Source: Ecology - nature.com
