Forecasts/projections for cherry blossom timing
Cherry trees bloom in the spring and then enter dormancy in autumn. The dormant period ends when a certain number of cold days have passed. The tree, sensing that winter is ending, begins growing buds, which grow faster when temperatures are warmer. Thus, lower winter temperatures and higher spring temperatures both contribute to earlier blossom dates.
The Japanese cherry blossom forecasts are based on the "DTS" method of Aono and Moriya (2003).
In this model, endodormancy ends
Aono & Moriya Table 2 shows these values as between day of year 30 and 50, which is somewhere in February.
After dormancy is over, the buds grow at a standard rate per day, when the air temperature is at a standard value. A day at a higher temperature constitutes more than one day transformed to standard temperature (DTS), and a colder day constitues less than one DTS, according to the Arrhenius equation:
where
Buds blossom when a certain, fixed number of DTS have elapsed since
The method of Aono & Moritz is ad hoc: the model does not reflect the underlying biological mechanism, and the use of
I use a different approach. Biological data suggests that, to exit endodormancy, cherry trees need to be below a certain threshold temperature, somewhere between 32 and 50
I assume that trees exit endodormancy after
I only consider a single location, so
For the purposes of this model, I interpolate
The Arrhenius equation is approximately linear over the relevant range, so I approximate:
where
The full bloom date
- Data
- Each season is considered independent of other seasons
-
$T(t)$ : temperature on day$t$ , in$\degree\text{C}$ -
$B$ : day of full flowering (this is the forecast target)
- Model parameters
-
$T_C$ : threshold temperature required to produce chill -
$C^\star$ : threshold number of chill units (i.e., days under threshold temperature) -
$A$ : DTS at 0$\degree\text{C}$ . -
$B$ : additional DTS per degree Celcius -
$\mathrm{DTS}^\star$ : threshold number of DTS before full bloom
-
- Priors and observation processes
-
$T_C \sim \mathrm{Unif}(-2.5, 5.0)$ , measured in$\degree$ C -
$C^\star \sim \mathrm{Unif}(5, 30)$ , measured in days below$T_C$ $A \sim \mathrm{Norm}(0.24, 0.1)$ $B \sim \mathrm{Norm}(0.04, 0.01)$ $\mathrm{DTS}^\star \sim \mathrm{Unif}(10, 100)$ -
$\varepsilon \sim \mathrm{Unif}(-7, 7)$ , observation error on$t_F$ , in days (i.e.,$t_F = \hat{t}_F + \varepsilon$ )
-
Aono & Moriya use an empirical approach, concluding that the end of endodormancy (typically in February) depends on the average temperature over January, February, and March. This is not causally consistent (March temperatures cannot affect February biology). Modeling the actual process as understood from biology is not particularly complicated.
More sophisticated forecasts in Japan account for weather data as well as observational data from sentinel trees.
E.g., another modeler used an empirical approach, concluding that blooms occurs after 400 or 600 cumulative daily degrees during spring.
There are some accounts from machine learning approach concluding that temperature is the main driver of blossom date, not precipitation or sunlight.
- EPA
- Cherry blossom dates and data file
- Historical temperatures for DC: https://www.ncdc.noaa.gov/cdo-web/datasets/GHCND/stations/GHCND:USW00013743/detail
- NPS
- https://www.nps.gov/subjects/cherryblossom/bloom-watch.htm
- Stage dates since 2004 (and two earlier peak bloom dates)
- Japan Meteorological Agency
- Aono et al.
- [https://www.ncei.noaa.gov/access/paleo-search/study/26430]
- 1200 years of peak bloom dates and mean March temperatures
- Kyoto
scrape_nps.ipynb
produces:data/nps.csv
: stage datesdata/nps_stages.csv
: order of the stages
scrape_aono.ipynb
producesdata/aono.csv
scrape_epa.ipynb
producesdata/epa.csv