Phenocam Forecast for NEFI summer course 2022

Team:

• Eric Scott
• Yiluan Song
• Ross Alexander
• Jussi Mäkinen

Background

Phenocams take automated daily photos of sites like this:

Then photos are converted to values of greenness and redness. These data can be used to forecast date of spring leaf-out or fall color change.

Challenge

• Predict greenness 35 days ahead from the current day (fixed to early February 2022)
• Observations start 2016

18 study sites in total

• 10 deciduous, 6 grassland, 2 shrubland

#> Warning: Removed 11245 rows containing missing values (geom_point).
#> Warning: Removed 342 rows containing missing values (geom_point).
#> Warning: Removed 2 row(s) containing missing values (geom_path).

(More data exploration)

Forecasting uncertainty

• model uncertainty
• climate forecast uncertainty (different forecast ensembles)

Problems

• account for the between-sites variation in temporal patterns and in response to climate
• easy access only to most recent climate data, constrained analysis to 2020->

Model structure

• https://www.sciencedirect.com/science/article/abs/pii/S0034425720303266

Data model

Process model

Priors

JAGS code:

model{
  
  for (s in 1:ns) {
    #### Data Model
    for(t in 1:nt){
      gcc[t, s] ~ dnorm(x[t, s],tau_obs_gcc)
      evi[t, s] ~ dnorm(x[t, s],tau_obs_evi)
    }
    
    #### Process Model
    for(t in 2:nt){
      x[t, s]~dnorm(x[t-1, s],tau_add)
    }
  }
  
  
  #### Priors
  for (s in 1:ns) {
     x[1, s] ~ dnorm(x_ic[s],tau_ic[s])
  }
  tau_obs_gcc ~ dgamma(a_obs_gcc,r_obs_gcc)
  tau_obs_evi ~ dgamma(a_obs_evi,r_obs_evi)
  tau_add ~ dgamma(a_add,r_add)
}

Forecasts

1. Prepare new data for assimilation
2. Load posterior as prior/ initialize uninformative prior
3. Set initial conditions
4. Configure model
5. Fit model (and forecast)
6. Model assessment
7. Summarize posteriors with hyperparameters (save hyperparameters)
8. Combine previous data with forecast (save data)
9. Visualize (save plots)

Some examples

• 2020-11-24
• 2021-07-22
• 2022-05-18

targets workflow:

graph LR
  subgraph legend
    x7420bd9270f8d27d([""Up to date""]):::uptodate --- x5b3426b4c7fa7dbc([""Started""]):::started
    x5b3426b4c7fa7dbc([""Started""]):::started --- xbf4603d6c2c2ad6b([""Stem""]):::none
  end
  subgraph Graph
    x9d9338876342a883(["all_dat"]):::uptodate --> x16fdb873ff498824(["all_dat_ano"]):::uptodate
    x68dd683e0472743b(["mindate"]):::uptodate --> x16fdb873ff498824(["all_dat_ano"]):::uptodate
    xb8d8de52ba56a7bb(["gcc_dat_file"]):::uptodate --> x250fae475e168023(["gcc_dat"]):::uptodate
    x37a4b6e78faf3120(["noaa_dat_file"]):::uptodate --> xb9d1c1bbc12ef44d(["noaa_dat"]):::uptodate
    x8fa3f0bfe3afe1bc(["RandomWalk"]):::uptodate --> x793b57f9be3e25d5(["README"]):::started
    x32cef2290a81c584(["ts_plot"]):::uptodate --> x793b57f9be3e25d5(["README"]):::started
    x9d9338876342a883(["all_dat"]):::uptodate --> xab6329fc3aa0e7b8(["forecasts"]):::uptodate
    x74653413816894b0(["batch"]):::uptodate --> xab6329fc3aa0e7b8(["forecasts"]):::uptodate
    x8448797b328e6352(["date_list"]):::uptodate --> xab6329fc3aa0e7b8(["forecasts"]):::uptodate
    x68dd683e0472743b(["mindate"]):::uptodate --> xab6329fc3aa0e7b8(["forecasts"]):::uptodate
    x8fa3f0bfe3afe1bc(["RandomWalk"]):::uptodate --> xab6329fc3aa0e7b8(["forecasts"]):::uptodate
    x18f6cc99ecb95617(["hls_df_file"]):::uptodate --> x64431175705194ee(["hls_df_proc"]):::uptodate
    x74653413816894b0(["batch"]):::uptodate --> x8448797b328e6352(["date_list"]):::uptodate
    x250fae475e168023(["gcc_dat"]):::uptodate --> x8448797b328e6352(["date_list"]):::uptodate
    x68dd683e0472743b(["mindate"]):::uptodate --> x8448797b328e6352(["date_list"]):::uptodate
    xb9d1c1bbc12ef44d(["noaa_dat"]):::uptodate --> x68dd683e0472743b(["mindate"]):::uptodate
    x9d9338876342a883(["all_dat"]):::uptodate --> x32cef2290a81c584(["ts_plot"]):::uptodate
    x68dd683e0472743b(["mindate"]):::uptodate --> x32cef2290a81c584(["ts_plot"]):::uptodate
    x16fdb873ff498824(["all_dat_ano"]):::uptodate --> xdae95d0f164d35b8(["anom_plot"]):::uptodate
    x68dd683e0472743b(["mindate"]):::uptodate --> xdae95d0f164d35b8(["anom_plot"]):::uptodate
    x250fae475e168023(["gcc_dat"]):::uptodate --> x9d9338876342a883(["all_dat"]):::uptodate
    x64431175705194ee(["hls_df_proc"]):::uptodate --> x9d9338876342a883(["all_dat"]):::uptodate
    xb9d1c1bbc12ef44d(["noaa_dat"]):::uptodate --> x9d9338876342a883(["all_dat"]):::uptodate
    xcd447f216a0c85c5(["EDA"]):::uptodate --> xcd447f216a0c85c5(["EDA"]):::uptodate
  end
  classDef uptodate stroke:#000000,color:#ffffff,fill:#354823;
  classDef started stroke:#000000,color:#000000,fill:#DC863B;
  classDef none stroke:#000000,color:#000000,fill:#94a4ac;
  linkStyle 0 stroke-width:0px;
  linkStyle 1 stroke-width:0px;
  linkStyle 25 stroke-width:0px;

Resources for Challenge

• challenge docs
• phenocam

gcc data is here:

gcc_dat <- 
  readr::read_csv(
    "https://data.ecoforecast.org/targets/phenology/phenology-targets.csv.gz",
    guess_max = 1e6
  )

site metadata is here:

site_data <- 
  readr::read_csv(
    "https://raw.githubusercontent.com/eco4cast/neon4cast-phenology/master/Phenology_NEON_Field_Site_Metadata_20210928.csv"
    )

Repo structure

• data/ put raw data here
• R/ put R functions to be source()ed here
• docs/ put .Rmd files to be rendered here