(C) Daily Yonder - Keep it Rural
This story was originally published by Daily Yonder - Keep it Rural and is unaltered.
. . . . . . . . . .
The effects of climate change on the timing of peak fall foliage in Acadia National Park [1]
['Spera', 'Stephanie A.', 'Sspera Richmond.Edu', 'Department Of Geography', 'Environment', 'Sustainability', 'University Of Richmond', 'Richmond', 'The Schoodic Institute', 'Winter Harbor']
Date: 2023-09-17
Acadia National Park, Mount Desert Island
Mount Desert Island is home to over 85% of Acadia National Park. We focus our analysis on Mount Desert Island (MDI) for two reasons. Firstly, a majority of the deciduous trees in Acadia are found on MDI (Fig. 3), where The Great Fire of 1947 burned over 40 km2 of spruce and fir trees—including nearly one-third of the total park area on Mount Desert Island, and maples, birch trees, and aspens have grown in their place (NPS 2020). Secondly, in 2021, over 90% of the Acadia visitors remained on MDI; less than 1% visited Isle au Haut and the rest visited the Schoodic Peninsula (Broom 2021). We also extend our time period of analysis to the pre-satellite data era from 1950 to 2021 to facilitate robust statistical analyses.
Fig. 3 Predominant land use and vegetation types across Mount Desert Island and Acadia National Park. Data were created in 2003 as part of a Vegetation Mapping Inventory Project in collaboration with the National Park Service, USGS, Maine Natural Areas Program, and The Nature Conservancy (NPS Project 1047691 2003). Deciduous forest land is highlighted in orange Full size image
Climate data and analysis
Because there is no consistent long-term observational record on MDI (see SI for MDI GHCNd station information), we downloaded 9 km resolution global land-based reanalysis ERA5-Land hourly precipitation and 2 m air temperature data from January 1, 1950–December 31, 2021 (Muñoz-Sabater et al. 2021), and derived daily total precipitation, and minimum and maximum temperatures. ERA5-Land data is improved from the European Centre for Medium-Range Weather Forecasts’ (ECMWF) fifth version of European ReAnalysis (ERA5) as it focuses solely on land, includes an elevation correction, and offers higher-resolution data (Muñoz-Sabater et al. 2021). Comparisons with GHCNd data from stations on MDI demonstrate that ERA5-Land data provide robust representations of weather and climate dynamics across the study area (SFigs. 1, 2, 3).
We used these daily data both to analyze overall trends in seasonal climate in Acadia National Park between 1950 and 2021 and to derive independent variables for our statistical modeling of the timing of peak fall foliage, as described in the “Statistical Analyses” section below. We first calculated seasonal averages for each variable across Mount Desert Island, i.e., average winter (December, January, February), spring (March, April, May), summer (June, July, August), and fall (September, October, November) minimum temperature, maximum temperature, and total precipitation.
We also determined 95th and 99th percentile temperature and precipitation thresholds over our entire time period. For each smaller time period of interest, we calculated meteorological variables including “warm days”—defined as the number of days warmer than the 95th percentile maximum temperature threshold; “hot days”—defined as the number of days warmer than the 99th percentile maximum temperature threshold; “warm nights”—number of nights warmer than the 95th minimum temperature threshold; “hot nights”—number of nights warmer than the 99th minimum temperature threshold; “cool nights”—number of nights between the 54% and 73% percentile (~ 4.5–10 °C); “rain days”—number of days with more precipitation than the 50th percentile of daily precipitation; “pour days”—number of days with precipitation greater than the 95th percentile; “downpour days”—number of days with precipitation greater than the 99th percentile; and “droughts”—number of periods with more than seven consecutive days of no precipitation (See Supplementary Table 1 for a list of all variables). We note that with our “cool/warm/hot nights” variables, we did not subset the hourly data to nightly hours and thus assume that the daily minimum temperature represents the overnight low. Occasionally, the daily minimum temperature occurs during the day, like when a strong cold front moves in, but on average, daily minimum temperatures provide reasonable approximation of overnight temperatures.
All statistical tests were run using R v4.2.1. We ran Mann–Kendall (McLeod 2022) tests to determine if there were statistically significant trends in the aforementioned values over each time period of interest. The output of this test also provided Kendall’s tau correlation coefficient, which determines the strength of the trend (values close to − 1 suggest a strong decreasing trend and values close to 1 suggest a strong increasing trend). We calculated the median rate of change of each significant trend using the Sen’s slope estimator (Pohlert 2020).
We also used the coefficients from our final ERA5-Land based model to predict 2022 fall foliage timing, first using January 1, 2022–October 31, 2022 ERA5-Land data as the input independent variables, and then using Global Historical Climatology Network daily (GHCNd) data. Because there is no station on Mount Desert Island that collects both temperature and precipitation, we downloaded daily temperature data from the McFarland Station (USR0000MMCF) and downloaded and averaged daily precipitation data from the two precipitation stations in Southwest Harbor (Southwest Harbor 2.6 SE—US1MEHN0003, Southwest Harbor 0.9 NW—US1MEHN0064).
Remotely-sensed data and analysis
The timing of prolonged events, such as the duration and stages of fall leaf senescence and coloration, can be difficult to define, and is particularly tricky across Mount Desert Island where over half of the vegetation is classified as evergreen forest land (Fig. 3). A majority of institutions and states use percentage of leaves that have changed color to categorize various stages of fall foliage: the Maine Department of Agriculture, Conservation, and Forestry, which releases weekly reports in the fall, has defined five periods of fall foliage: very low—0–10% of leaves have changed color; low—10–30% of leaves have changed color, moderate—30–50%, high—50–70%, and peak—70–100%; weather.com has four defined categories of fall foliage: no change—0–10%; patchy—10–50%, near peak—50–75%, and peak, 75–100%; and the long-term ecological research (LTER) site at Harvard Forest records the percentage of leaf fall and leaf color every three to seven days. The United States Forest Service, however, uses five descriptive categorical values (0–4) to monitor fall foliage in Hubbard Brook, New Hampshire, e.g., “0: Most leaves fallen” and “2: Most leaves becoming yellow or red, and a few fallen leaves.” And, when discussing with park visitors and local residents about what is ‘peak’ fall foliage, one can sum up their answers as “I know it when I see it.” Thus, not only is monitoring fall foliage using remotely-sensed data difficult, but so too is validating the remotely sensed results in areas without long-term monitoring stations.
We adjusted the normalized brownness index (NBI) methodology developed by Zhang and Goldberg (2011). To calculate NBI, Moderate Resolution Imaging Spectroradiometer (MODIS) data is used to calculate annual normalized difference vegetation index (NDVI) timeseries. Then, a sigmoidal model is applied to the senescent phase of the smoothed NDVI timeseries (Zhang and Goldberg 2011). Then, using principles of spectral mixture modeling, fractions of ‘green material’ and ‘brown material’ are isolated, and the fraction of brown material parameter, i.e., the normalized brownness index, can then be used to determine fall foliage phases corresponding to percentage of leaf coloration and leaf fall (Zhang and Goldberg 2011). See Zhang and Goldberg (2011) for a complete theoretical background and methodology.
The two alterations we made to the algorithm were related to the input data and the annual time series data. Following Zhang and Goldberg (2011), we downloaded projected, mosaicked, and subsetted 500 m resolution MODIS MC43A4 Version 6 Nadir bidirectional reflectance distribution function (BDRF)-adjust reflectance (NBAR) data between February 5, 2000 (first data available) and December 31, 2021 over our study area using the USGS application for extracting and exploring analysis ready samples (AρρEEARS) tool. However, Zhang and Goldberg (2011) used a previous version of the dataset with a temporal resolution of 8 days, whereas we used Version 6 which has a daily temporal resolution. Also, instead of calculating annual NDVI time series to back out the NBI, we calculated, smoothed, and fit a sigmoid function to annual enhanced vegetation index (EVI) time series, because EVI better corrects for atmospheric conditions, reduces noise and is more sensitive to changes in vegetation cover (Huete et al. 1999).
We compare the results of the remotely sensed data to an archival historical analysis of peak fall foliage described in the “Historical Analysis” section below as another means of validating the archival work, but we note that we do not include these remotely sensed data in the statistical analyses described in "Statistical analyses" section. Moreover, calculating NBI works best when a pixel is predominantly deciduous forest. As highlighted by Fig. 3, although we ran our algorithm over each of the 2719 MODIS pixels that cover Mount Desert Island, only 104 pixels were comprised of more than 50% of deciduous or deciduous and mixed forest cover, only 24 pixels were comprised of more than 50% of deciduous forest cover, and only three pixels were comprised of more than 75% of deciduous forest cover. Thus, we focused our remotely sensed data analysis on the 104 pixels that contain at least 50% deciduous or deciduous and mixed forest cover (approximately 26 km2).
Historical analysis
Because we both wanted to extend our time series of fall foliage to the pre-MODIS data era (pre-2001), and we required data to validate our remotely sensed analysis, we performed an archival analysis to determine the timing of peak fall foliage. Much effort was made to contact individuals, businesses, and scientific organizations in and around Mount Desert Island to collect historical information on fall foliage. Moreover, although the Maine Department of Agricultural, Conservation and Forestry (DCAF) has been keeping records of the timing of fall foliage across Maine for decades, only recently have they reported these weekly records online, and a majority of the paper records were thrown out (M. Macaluso, personal communication). Thus, we used National Park Service Acadia Monthly Narrative Reports, newspaper archives, fall foliage reports, and citizen science submitted photos to reconstruct a time-series of peak fall foliage dating back to 1951 (Table 1). Screenshots of a subset of the reports, newspaper clips, and photos can be found in the Supplementary Information (SI). We recognize that ‘peak’ fall foliage can be a multi-day event, and thus note that that choosing one day as ‘peak’ for each year during which we have data is not ideal. However, we do believe we have derived the best representation of the timing of peak fall foliage over our study period.
Table 1 Sources of timing of peak fall foliage from our archival research Full size table
Similar to the climate data analysis above, we conducted a Mann–Kendall test and calculated the Sen’s slope estimator to determine if there were statistically significant trends in the timing of peak fall foliage over our time period.
Statistical analyses
We used multivariate linear regressions to determine the broad climatic drivers of the timing of peak fall foliage across Mount Desert Island.
Correlation analysis
We derived over 130 climatic variables between 1950 and 2021 using the same metrics as highlighted in the ‘Climate Data and Analysis Section,’ but broken down into even smaller time periods including, seasonal, monthly, and weekly time scales (Supplementary Table 2). Precipitation variables included our aforementioned total precipitation, number of ‘pour days,’ number of ‘downpour days,’ and number of ‘drought periods’ within each timescale of interest. We categorized temperature variables in three ways: (1) we calculated mean temperatures within each time period of interest (week, month, season); (2) we calculated the average minimum and maximum temperature within each time period of interest; and (3) we calculated the number of warm nights, hot nights, cool nights, warm days, and hot days within each time period of interest. Following Archetti et al. (2013), we then determined the correlation coefficient between each derived variable and peak fall foliage, and we identified those that were significant at a p-value ≤ 0.10 to then use as independent variables in our regression modeling.
Regression modeling and model selection
We used a variety of variable selection techniques to derive models of the timing of peak fall foliage in Acadia National Park. We performed variable selection using best subsets (Miller 2020), best subsets with a validation set, k-fold cross-validation (Davison and Hinkley 1997; Canty and Ripley 2021), and lasso selection approaches. The best subsets technique identifies the set of variables that maximizes an objective indicator of model fit, including the residual sum of squares (RSS), adjusted R2, Bayesian Information Criterion (BIC) and Mallows C p statistic values (Beale et al. 1967; Hocking and Leslie 1967). For our purposes, we used adjusted R2 (“Best Subsets A” below) and BIC (“Best Subsets B” below). The validation set approach splits the dataset into a training and validation set and computes the MSE of each best subsets selection model to determine which has the lowest MSE. The full dataset is then used to compute the coefficients in the final model. The k-fold cross validation takes the validation set approach further, splitting the data into k parts (here we used k = 10), or folds, and using k−1 parts to train the sample and the remaining partition to validate it. The prediction errors are then averaged, and the model with the lowest error is selected. Lastly, lasso (least absolute shrinkage and selection operation) performs variable selection by penalizing complex models while still minimizing the sum of squares to determine the best model (Tibshirani 1996).
Variables from each temperature category described above were never combined in our variable selection process, thus, each variable selection process was run three times using each category of temperature variables. We ensured each model’s selected variables did not exhibit multicollinearity by calculating variance inflation factor (VIF) values and removing selected variables if the VIF was greater than 5. Then, to choose the best-fit model of those derived from our variable selection processes highlighted above, we compared Akaike information criterion (AIC) and BIC values calculated using the full dataset, and the root-mean-squared-error (RMSE) and adjusted R2 values, which were calculated using leave-one-out-cross validation.
We used our best-fit model to predict the timing of peak fall foliage in 2022 using both GHCN-d and ERA5-Land 2022 data as independent input datasets and with the future climate projection analysis described below.
Future climate scenarios
To determine how the timing of peak fall foliage might be expected to change in the future, we compared climate projection data from three Shared Socioeconomic Pathways: SSP1-2.6, SSP2-4.5, and SSP5-8.5. These SSPs consider socioeconomic factors, and how the world might address both climate change adaptation and mitigation. SSP1-2.6 is the “Sustainability-taking the green road” pathway, where global temperatures stay below an excess 2 °C of warming and there are low challenges to mitigation and adaptation; SSP2-4.5-is considered the “Middle of the road” pathway with medium challenges to mitigation and adaptation, and SSP5 is the “Fossil-fuel development” pathway with high challenges to climate adaptation and mitigation (O’Neill et al. 2017). We downloaded coupled model intercomparison project (CMIP6) (CESM2) community earth system model daily maximum near-surface air temperature, daily minimum near-surface air temperature, and total daily precipitation data from January 1, 2022-December 31, 2060 (O’Neill 2016; Danabasolgu et al. 2020).
[END]
---
[1] Url:
https://link.springer.com/article/10.1007/s10980-023-01703-0
Published and (C) by Daily Yonder - Keep it Rural
Content appears here under this condition or license: Creative Commons CC BY-ND 4.0 International.
via Magical.Fish Gopher News Feeds:
gopher://magical.fish/1/feeds/news/dailyyonder/
