PEG Models used in the EFI-NEON Forecasting Challenge

The GenoPhenoEnvo team submitted forecasts of forest phenology inferred from webcams to the EFI Spring 2021, Fall 2021, and Spring 2022 Challenges. Results for submitted forecasts can be viewed on the EFI-NEON Ecological Forecasting Challenge Dashboard, here, where you can explore results such as:

Background

"The Ecological Forecasting Initiative is a grassroots consortium aimed at building and supporting an interdisciplinary community of practice around near-term (daily to decadal) ecological forecasts." - ecoforecast.org/about{.uri}

The EFI-NEON Forecasting Challenge

The National Science Foundation funded Ecological Forecasting Initiative Research Coordination Network (EFI-RCN) is hosting a NEON Ecological Forecast Challenge with the goal to create a community of practice that builds capacity for ecological forecasting by leveraging NEON data products. - projects.ecoforecast.org/neon4cast-docs

This repository contains models used in the Phenology challenge, described in more detail the "Phenology" chapter of the EFI-NEON Ecological Forecasting Challenge documentation.

EFI and the Phenology challenge are best described in the links above. In addition, you may be interested in the following resources:

• The EFI YouTube channel, including the Phenology challenge description and an overview of NEON data streams.

• Publications by Andrew Richardson et al. on the phenocams:

  • Richardson, A., Hufkens, K., Milliman, T. et al. Tracking vegetation phenology across diverse North American biomes using PhenoCam imagery. Sci Data 5, 180028 (2018). https://doi.org/10.1038/sdata.2018.28

  • Richardson, A.D. (2019), Tracking seasonal rhythms of plants in diverse ecosystems with digital camera imagery. New Phytol, 222: 1742-1750. https://doi.org/10.1111/nph.15591

Repository Contents

A Simple Model (PEG)

Lead: David LeBauer

The original aim of having a 'simple model' was work out the mechanisms of the forecast challenge. We call it 'simple' because these models have a single input - the historical time series of the variable (gcc or rcc) from 2016 through day t-1 to predict values of the variable through day t+1:t+180.

Although it started as a 'simple' moving window prediction, we later implemented an seasonal plus exponential smoothing model using the R forecast package.

Inputs and outputs were the same for both models. Neither model used gcc or rcc.

• Inputs: historical time series of gcc or rcc from NEON Phenocam sites.

• Outputs: daily, 180 day forecasts of gcc and rcc

Code:

• simplefolder:
  • ets_forecast.R An exponential smoothing model with seasonality using the forecast package in R. The original moving window model is in the comments

Moving Window Mean Through 2021-03-17

This model used a rolling mean around previous years' dates to predict future. The forecast for day d was a rolling average of the 8 days before and after.

Seasonal + Exponential Smoothing 2021-03-17 to the end of the year

This is not actually a simple model, it is in fact a very sophisticated Seasonally-adjusted exponential smoothing state-space model. In this case, it is only simple in that it is both univariate and easy to implement using the forecast R package (Hyndman & Khandakar, 2008) following the clear explanations provided in the Forecasting Principles and Practice text (Hyndman & Athanasopoulos, 2018).

Hyndman, R.J., & Athanasopoulos, G. (2018) Forecasting: principles and practice, 2nd edition, OTexts: Melbourne, Australia. OTexts.com/fpp2. Accessed on <current date>.

Hyndman RJ, Khandakar Y (2008). "Automatic time series forecasting: the forecast package for R." _Journal of Statistical Software_, *26*(3), 1-22. doi: 10.18637/jss.v027.i03 (URL:https://doi.org/10.18637/jss.v027.i03)..)

Ideas for future improvement / additional models

• Holts-Winter seasonality model
• Fit ARIMA model using the forecast::auto.arima function.

Machine Learning Models (Team PEG_RFR, PEG_FUSION)

Leads: Arun Ross and Debashmita Pal

• ML folder
  • See details in that folder's README